Customer service question and answer processing method, device, equipment and medium based on multi-agent large model

Abstract

This application discloses a customer service question-and-answer processing method, apparatus, device, and medium based on a multi-agent large model, comprising: obtaining conversation data between a customer service representative and a user, preprocessing the conversation data, and generating a preprocessed conversation text; performing semantic analysis on the conversation text to identify user questions and customer service answers, and generating question-and-answer pairs containing the problem resolution status; determining whether the customer service answer perfectly resolves the user's problem, and generating a set of question-and-answer pairs marked as perfectly resolved; classifying and arranging the set of question-and-answer pairs marked as perfectly resolved to summarize similar question-and-answer pairs and extract structured knowledge items; storing similar question-and-answer pairs in a professional knowledge base based on the structured knowledge items, and updating the professional knowledge base; and providing corresponding answers to the customer service representative based on the professional knowledge base in response to receiving actual user business questions. Through the collaborative work of multiple agents, the efficiency and accuracy of customer service responses to user questions are improved.

Description

Technical Field

[0001] The present application relates to the field of large model applications, and in particular to a customer service question and answer processing method, device, equipment and medium based on a multi-agent large model. Background Art

[0002] In today's digital age, customer service systems have become a crucial bridge between businesses and users. However, existing customer service systems present numerous challenges in practical applications. First, traditional customer service systems have limited understanding of user questions and often fail to accurately grasp the user's true intent, resulting in significant deviations in responses. Second, the quality of customer service responses varies widely, lacking unified standards and specifications, making it difficult to guarantee high-quality answers. Furthermore, the knowledge base construction and update mechanisms are inadequate, failing to promptly reflect business changes and user needs, resulting in outdated knowledge base content and a lack of effective support for customer service. These issues severely impact user experience and customer service efficiency, hindering improvements in service quality and competitiveness for businesses. Summary of the Invention

[0003] The purpose of this application is to provide a customer service question and answer processing method, device, equipment and medium based on a multi-agent large model, so as to at least solve the problem that the current traditional customer service system method cannot promptly reflect business changes and user needs and provide users with accurate and efficient answers.

[0004] To solve the above technical problems, this application provides a customer service question-answering processing method based on a multi-agent large model, including:

[0005] Using the recognition agent to obtain the conversation data between the customer service and the user, preprocessing the conversation data to generate a preprocessed conversation text;

[0006] Using the recognition agent to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status;

[0007] Using the judging agent to perform a multi-dimensional evaluation of the question-answer pairs based on preset judging criteria, determine whether the customer service answer perfectly solves the user's problem, and generate a set of question-answer pairs marked as perfectly solved;

[0008] Using the constructed agent to classify and organize the set of question-answer pairs marked as perfectly solved, so as to summarize similar question-answer pairs and extract structured knowledge items;

[0009] Using the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base;

[0010] The response agent is used to respond to the actual user business question received and provide the customer service with a corresponding answer based on the professional knowledge base.

[0011] Optionally, the use of the recognition agent to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status, including:

[0012] Using a large model fine-tuned in the customer service field, the pre-processed conversation text is subjected to intent recognition and semantic segmentation to distinguish the boundary between user questions and customer service responses.

[0013] Match the user's question with the customer service answer based on contextual analysis;

[0014] The problem solving status is determined through a logical reasoning model. If the customer service answer covers the requirements of the user's problem and there is no contradiction, it is marked as solved, and a question-answer pair containing the problem solving status is generated.

[0015] Optionally, the preset evaluation criteria include at least one of the following dimensions:

[0016] Accuracy evaluation: Use a large model to verify the consistency of the customer service answer with the domain knowledge base;

[0017] Completeness evaluation: Analyze whether the customer service response covers all sub-requirements of the user's question;

[0018] Timeliness evaluation: Compare customer service response time with preset thresholds;

[0019] User satisfaction prediction: Outputs satisfaction scores based on the conversation sentiment analysis model.

[0020] Optionally, the construction agent is used to classify and organize the set of question-answer pairs marked as perfectly solved to summarize similar question-answer pairs and extract structured knowledge items, including:

[0021] Perform semantic vectorization on the question-answer pairs marked as perfectly solved, and use clustering algorithms to divide similar question-answer pairs;

[0022] Based on the text summarization capability of the large model, the common answer framework and differences are extracted from each group of similar question and answer pairs;

[0023] The common answer framework is merged with the difference points to generate structured knowledge entries covering multiple scenarios.

[0024] Optionally, after merging the common answer framework with the difference points to generate structured knowledge items covering multiple scenarios, the method further includes:

[0025] Modeling the dependency relationships between the structured knowledge items through a graph neural network (GNN) to generate a knowledge graph;

[0026] The node weights of the knowledge graph are dynamically adjusted according to the user query frequency to optimize the retrieval priority.

[0027] Optionally, the step of using the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base further includes:

[0028] Adding a version identifier and an effective timestamp to the newly generated structured knowledge entry;

[0029] Regularly detecting the reference frequency of each structured knowledge item in the professional knowledge base according to the version identifier and the effective timestamp, and archiving the items with a reference frequency lower than a preset value into a low-frequency item history library;

[0030] The priority retrieval rule is updated based on the reference frequency.

[0031] Optionally, the preprocessing of the conversation data includes:

[0032] Denoising the conversation data, deleting irrelevant symbols and non-text information, and obtaining a conversation text;

[0033] Dividing the conversation text into sentences according to semantic coherence and converting the sentences into a unified coding format;

[0034] Based on the dialogue text in the unified coding format, scene tags of the dialogue text are extracted, and the scene tags are associated with the dialogue time and user identity information to generate pre-processed dialogue text.

[0035] To solve the above technical problems, the present application also provides a customer service question-answering processing device based on a multi-agent large model, comprising:

[0036] The first recognition module is used to use the recognition agent to obtain the conversation data between the customer service and the user, pre-process the conversation data, and generate a pre-processed conversation text;

[0037] A second recognition module is used to use the recognition agent to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status;

[0038] The question-answer evaluation module is used to use the evaluation agent to perform a multi-dimensional evaluation of the question-answer pairs based on preset evaluation criteria, determine whether the customer service answer perfectly solves the user's problem, and generate a set of question-answer pairs marked as perfectly solved;

[0039] The first construction module is used to classify and organize the set of question-answer pairs marked as perfectly solved using a construction agent, so as to summarize similar question-answer pairs and extract structured knowledge items;

[0040] A second construction module is configured to utilize the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base;

[0041] The business response module is used to use the response agent to respond to actual user business questions received and provide corresponding answers to customer service based on the professional knowledge base.

[0042] In order to solve the above technical problems, the present application also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the above-mentioned customer service question and answer processing method based on the multi-agent large model.

[0043] In order to solve the above technical problems, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the above-mentioned customer service question and answer processing method based on the multi-agent large model.

[0044] The beneficial effects of the embodiments created by the present application are as follows: by using the recognition agent to obtain the conversation data between the customer service and the user, the conversation data is pre-processed to generate a pre-processed conversation text; using the recognition agent to perform semantic analysis on the pre-processed conversation text, identify the user question and the customer service answer, and generate a question-answer pair containing the problem-solving status; using the judgment agent based on preset judgment criteria to perform a multi-dimensional evaluation of the question-answer pair, determine whether the customer service answer perfectly solves the user problem, and generate a set of question-answer pairs marked as perfectly solved; using the construction agent to classify and organize the set of question-answer pairs marked as perfectly solved to summarize similar question-answer pairs and refine structured knowledge items; using the construction agent to store the summarized similar question-answer pairs in a professional knowledge base based on the structured knowledge items to update the professional knowledge base; using the response agent to respond to actual user business questions received and provide corresponding answers to the customer service based on the professional knowledge base. Based on the collaborative work of multiple agents and the use of natural language processing technology based on large models, user questions can be accurately identified and understood. At the same time, the question and answer evaluation agent conducts a comprehensive evaluation of customer service answers, ensuring the high quality and consistency of the answers, and making up for the uneven answer quality in traditional customer service systems. In addition, the knowledge base is used to build an agent that can integrate high-quality question and answer pairs in real time and dynamically update the knowledge base to make its content more timely and practical, solving the problems of slow updates and outdated content in the existing knowledge base, improving the efficiency and accuracy of customer service answers to user questions, and significantly improving the intelligence level of the customer service system and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0046] Figure 1 This is a basic flow chart of a customer service question-and-answer processing method based on a multi-agent large model according to a specific embodiment of the present application;

[0047] Figure 2 This is a schematic diagram of the basic structure of a customer service question and answer processing device based on a multi-agent large model according to a specific embodiment of the present application;

[0048] Figure 3 This is a basic structural block diagram of a computer device according to a specific embodiment of the present application. DETAILED DESCRIPTION

[0049] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and are not to be construed as limiting the present application.

[0050] Those skilled in the art will understand that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of this application refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0051] Those skilled in the art will understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless specifically defined as such, will not be interpreted in an idealized or overly formal sense.

[0052] Those skilled in the art will appreciate that the term "terminal" as used herein includes both devices that are wireless signal receivers, i.e., devices that only have wireless signal receivers without transmission capabilities, and devices that have receiving and transmitting hardware capable of performing two-way communication over a two-way communication link. Such devices may include: cellular or other communication devices with single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) devices that may combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include a radio frequency receiver, a pager, Internet / Intranet access, a web browser, a notepad, a calendar, and / or a GPS (Global Positioning System) receiver; and conventional laptop and / or palmtop computers or other devices that have and / or include a radio frequency receiver. As used herein, a "terminal" can be portable, transportable, installed in a vehicle (air, sea, and / or land), or adapted and / or configured to operate locally and / or in a distributed manner at any other location on Earth and / or in space. A "terminal" as used herein can also refer to a communication terminal, an Internet access terminal, or a music / video playback terminal, such as a PDA, a mobile internet device (MID), and / or a mobile phone with music / video playback capabilities, as well as devices such as smart televisions and set-top boxes.

[0053] The hardware referred to by names such as "server", "client", and "service node" in this application is essentially an electronic device with capabilities equivalent to those of a personal computer. It is a hardware device that has the necessary components revealed by the von Neumann principle, such as a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device. Computer programs are stored in its memory, and the central processing unit loads the program stored in the external memory into the internal memory for execution, executes the instructions in the program, and interacts with the input and output devices to complete specific functions.

[0054] It should be noted that the concept of "server" referred to in this application can also be extended to server clusters. Based on the network deployment principles understood by those skilled in the art, the servers described should be logically divided. In physical space, these servers can be independent of each other but callable through interfaces, or integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method of this application.

[0055] Unless expressly specified, one or more technical features of the present application can be deployed on a server for implementation and accessed by a client through a remote call to obtain an online service interface provided by the server, or can be directly deployed and run on a client for implementation.

[0056] Unless explicitly specified, the AI ​​models referenced or may be referenced in this application can be deployed on a remote server and remotely called on the client, or can be deployed and directly called on a client with sufficient device capabilities. In some embodiments, when it runs on the client, its corresponding intelligence can be obtained through transfer learning to reduce the requirements for the client's hardware operating resources and avoid excessive occupation of the client's hardware operating resources.

[0057] Unless explicitly specified, the various data involved in this application can be stored remotely on a server or on a local terminal device, as long as they are suitable for being called by the technical solution of this application.

[0058] Those skilled in the art should be aware that although the various methods of this application are described based on the same concept and thus exhibit commonality, unless otherwise specified, these methods can be independently executed. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept. Therefore, concepts with the same expression, as well as concepts that are appropriately transformed for convenience despite different expression, should be understood as equivalent.

[0059] Unless expressly stated to be mutually exclusive, the various embodiments disclosed in this application may be cross-combined with the relevant technical features of the various embodiments to flexibly construct new embodiments, as long as such combination does not deviate from the creative spirit of this application and can meet the needs of the prior art or resolve certain deficiencies in the prior art. Those skilled in the art should be aware of such flexibility.

[0060] See also Figure 1 , Figure 1 Schematic diagram of the basic flow of the customer service question and answer processing method based on the multi-agent large model in this embodiment.

[0061] like Figure 1 As shown, including:

[0062] S1100: Using the recognition agent to obtain conversation data between the customer service representative and the user, preprocessing the conversation data to generate a preprocessed conversation text;

[0063] This embodiment can be applied to customer service scenarios in various fields, including finance, e-commerce, insurance, education, healthcare, law, and hotel management. In this embodiment, a customer service system is configured to provide customer service personnel with accurate, efficient, and professional question-and-answer services. The system is configured with multiple agents. An agent is an intelligent module or software component with specific functions that can automatically perform specific tasks or process data. Agents are typically based on artificial intelligence technologies, such as natural language processing and machine learning, and can efficiently complete complex data processing tasks. The system is configured with recognition agents, evaluation agents, construction agents, and response agents. The system first uses the recognition agent to obtain conversation data between customer service personnel and users, preprocesses the conversation data, and generates preprocessed conversation text. Specifically, the recognition agent collects conversation data between customer service personnel and users in real time through the customer service system interface. The conversation data includes conversation text, timestamps, and conversation context. The conversation data is then preprocessed by first removing noise from the conversation text, such as irrelevant symbols and repeated sentences. The conversation text is then segmented into independent sentence units. For example, the conversation text "Hello! I'd like to consult on how to use Product A. Thank you." will be segmented into three sentences: "Hello!", "I'd like to consult on how to use Product A," and "Thank you." This preprocessed conversation text is clearer and more standardized, providing high-quality input data for subsequent semantic analysis.

[0064] It should be noted that the system of this embodiment can also incorporate sentiment analysis of conversational text during preprocessing. Specifically, after collecting conversational data, the recognition agent uses a sentiment analysis algorithm to perform sentiment annotation on the text, determining the user's emotional state (e.g., satisfaction, dissatisfaction, neutrality, etc.). For example, if a user text message reads "This product is terrible and completely meets my needs!", the recognition agent will label it as "dissatisfied." This sentiment annotation not only provides richer contextual information for subsequent semantic analysis and enriches the information dimension of the text, but also provides multi-dimensional training data for subsequent systems, thereby providing more targeted services when solving user problems.

[0065] It should be noted that the system of this embodiment can also incorporate conversation topic identification during the preprocessing process. Specifically, after collecting conversation data, the recognition agent uses natural language processing technology to extract topics from the text and determine the main content of the conversation. For example, if a user asks, "I want to learn about the after-sales service policy for Product A," the recognition agent will extract the topic "after-sales service." Through topic identification, the preprocessed conversation text can more accurately reflect the core issues of concern to the user, providing a more accurate basis for subsequent question and answer recognition and knowledge base construction. It can also help customer service personnel quickly identify problems and improve service efficiency.

[0066] S1200: Using the recognition agent to perform semantic analysis on the pre-processed conversation text, identify the user's question and the customer service's answer, and generate a question-answer pair including the problem resolution status;

[0067] After the recognition agent acquires conversation data between a customer service representative and a user, preprocesses the conversation data, and generates a preprocessed conversation text, the recognition agent performs semantic analysis on the preprocessed conversation text, identifies the user's question and the customer service representative's response, and generates a question-answer pair indicating the problem's resolution status. Specifically, the recognition agent uses natural language processing techniques to understand the semantic structure of the text and extract key information from the user's question and the customer service representative's response. For example, in the preprocessed conversation text, "I would like to inquire about how to use Product A," the recognition agent can identify "Product A" and "How to use" as key information and mark it as the user's question. Simultaneously, in the customer service representative's response, "For how to use Product A, please refer to page 10 of the user manual," the recognition agent can extract "Page 10 of the user manual" as key information and mark it as the customer service representative's response. Through semantic matching and logical analysis, the recognition agent further determines whether the problem has been resolved and generates a question-answer pair indicating the problem's resolution status. For example, if the customer service representative's response directly resolves the user's problem, it is marked as "Resolved"; otherwise, it is marked as "Unresolved." Finally, the generated question-answer pair is passed to a subsequent agent for further processing.

[0068] It should be noted that after extracting key information from the user question and the customer service answer, the recognition agent of this embodiment uses a semantic matching algorithm based on the Vector Space Model (VSM) to calculate the similarity between the user question and the customer service answer. Specifically, the user question and the customer service answer are respectively converted into word vectors, and the cosine similarity between them is calculated. For example, the user question "How do I use product A?" and the customer service answer "Please refer to page 10 of the user manual for how to use product A" will be converted into corresponding word vectors. The calculated cosine similarity is high, indicating that the answer is highly relevant to the question, thereby judging that the problem has been solved. This optimized semantic matching algorithm can more accurately evaluate the degree of matching between questions and answers, and improve the accuracy of question-answer pair generation.

[0069] It should be noted that in addition to processing text data, the recognition agent in this embodiment can also process multimedia information related to the conversation, such as voice and images. For example, a user might ask a question via voice, "What is the color of Product A?" The recognition agent will convert the voice into text for semantic analysis, while also extracting information about the intonation and speaking speed from the voice to aid in the decision-making process. If the user uploads a photo of Product A, the recognition agent can use image recognition technology to extract the product's color information and match it with the user's question. This multimodal semantic analysis approach enables a more comprehensive understanding of user questions and provides more accurate responses, making it particularly suitable for customer service conversations in complex scenarios.

[0070] S1300: Using a judging agent to perform a multi-dimensional evaluation on the question-answer pair based on preset judging criteria to determine whether the customer service answer perfectly solves the user's problem, and generating a set of question-answer pairs marked as perfectly solved;

[0071] After the recognition agent performs semantic analysis on the pre-processed conversation text to identify the user question and the customer service response, generating a question-answer pair containing the problem resolution status, the evaluation agent then performs a multi-dimensional evaluation of the question-answer pair based on preset evaluation criteria to determine whether the customer service response perfectly resolves the user's problem, generating a set of question-answer pairs marked as perfectly resolved. Specifically, after receiving the question-answer pair containing the problem resolution status, the evaluation agent first performs a comprehensive evaluation of the question-answer pair based on preset evaluation criteria. These evaluation criteria include multiple dimensions such as the accuracy, completeness, timeliness, and user satisfaction of the response. For example, for a question-answer pair, the evaluation agent examines whether the customer service response accurately resolves the user's problem (accuracy), whether the response covers all key points of the user's problem (completeness), whether the response is provided within a reasonable timeframe (timeliness), and whether the user feedback on the response is positive (user satisfaction). Through a comprehensive evaluation of these dimensions, the evaluation agent determines whether the customer service response meets the "perfect resolution" criteria. If so, the question-answer pair is marked as perfectly resolved and added to the set of perfectly resolved question-answer pairs for subsequent knowledge base construction. For example, for the user question "What is the warranty period for Product A?" and the customer service representative's answer "The warranty period for Product A is one year, calculated from the date of purchase," the judging agent confirmed through multi-dimensional evaluation that the answer was accurate, complete, and timely, and therefore marked it as a perfect solution.

[0072] It should be noted that the evaluation agent of this embodiment not only relies on preset evaluation criteria when conducting multi-dimensional evaluation of question-answer pairs, but also refers to direct feedback from users. For example, after the conversation ends, the system will send a satisfaction questionnaire to the user, asking whether the user is satisfied with the customer service's answer. User feedback (such as "very satisfied", "satisfied", and "unsatisfied") will be quantified into a satisfaction score and incorporated into the comprehensive evaluation system. For example, if a user's satisfaction score for a question-answer pair is "very satisfied", even if the question-answer pair performs averagely in other dimensions (such as timeliness), the evaluation agent may still mark it as a perfect solution. This mechanism of introducing user feedback can more realistically reflect user needs and experience, further improving the accuracy and reliability of the evaluation.

[0073] It should be noted that in addition to the main judging agent, the system of this embodiment also introduces multiple auxiliary agents, each of which is responsible for evaluating a specific dimension of the question-answer pair. For example, one auxiliary agent is responsible for evaluating the accuracy of the answer, another is responsible for evaluating the completeness of the answer, a third is responsible for evaluating the timeliness of the answer, and so on. The evaluation results of these auxiliary agents are aggregated into the main judging agent, which then makes a comprehensive judgment. For example, if the majority of auxiliary agents believe that the question-answer pair meets the criteria for a perfect solution, the main judging agent will mark it as a perfect solution. This multi-agent collaborative evaluation mechanism can improve the transparency and reliability of the evaluation process, while also improving the overall performance of the system through division of labor and cooperation.

[0074] S1400, using the construction agent to classify and organize the set of question-answer pairs marked as perfectly solved, so as to summarize similar question-answer pairs and extract structured knowledge items;

[0075] The evaluation agent performs a multi-dimensional evaluation of the question-answer pairs based on preset evaluation criteria to determine whether the customer service response perfectly resolves the user's problem, generating a set of question-answer pairs labeled "perfectly resolved." The construction agent then categorizes and organizes this set of perfectly resolved question-answer pairs to summarize similar pairs and extract structured knowledge items. Specifically, after receiving the set of perfectly resolved question-answer pairs, the construction agent first classifies them according to preset classification rules. Classification rules are based on characteristics such as the question's topic, business domain, or keywords. For example, the question-answer pair "User question: What is the warranty period for Product A? Customer service answer: Product A's warranty period is one year, starting from the date of purchase" would be classified under the topic "After-sales service." Next, the construction agent uses natural language processing technology to perform semantic analysis on question-answer pairs in the same category, extracting key information and classifying similar pairs. For example, multiple question-answer pairs regarding Product A's warranty period would be summarized into a common question-answer template: "Product A's warranty period is one year, starting from the date of purchase." Finally, the agent is built to refine these summarized question-answer pairs into structured knowledge items, for example: "Question: What is the warranty period of product A? Answer: The warranty period of product A is one year, calculated from the date of purchase." These structured knowledge items will be stored in the professional knowledge base for subsequent query and use.

[0076] It should be noted that after receiving a set of perfectly solved question-and-answer pairs, the construction agent in this embodiment automatically classifies them using a clustering algorithm (such as K-Means or DBSCAN). The clustering algorithm groups the question-and-answer pairs based on their semantic similarity, with the question-and-answer pairs within each group being highly semantically similar. For example, all question-and-answer pairs regarding the usage of product A will be automatically clustered into a single category. The construction agent then analyzes each clustering result, extracts common information, and generates a universal question-and-answer template. This approach reduces manual intervention and improves induction efficiency, making it particularly suitable for large-scale data processing scenarios.

[0077] It should be noted that when the construction agent of this embodiment classifies and summarizes question-answer pairs, in addition to considering the semantic similarity of the text, it also extracts features such as the timestamp, user feedback, and conversation scenario of the question-answer pairs. For example, for question-answer pairs of the same question presented at different times or in different scenarios, the construction agent will analyze the differences in their responses and generate more targeted knowledge items. For example, the after-sales service policy for product A during a promotional period may be different from that during non-promotional periods. The construction agent will summarize and generate different knowledge items for each. This multi-dimensional feature extraction and summarization method can make the knowledge base more accurate and practical, better meeting the needs of users in different scenarios.

[0078] S1500: Using the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base;

[0079] After the construction agent categorizes and organizes the set of question-answer pairs marked as perfectly solved to summarize similar question-answer pairs and extract structured knowledge items, the construction agent then stores the summarized similar question-answer pairs in the professional knowledge base based on the structured knowledge items to update the professional knowledge base. In one embodiment, after completing the categorization, summarization, and extraction of the question-answer pairs, the construction agent generates structured knowledge items, such as "Question: What is the warranty period for Product A? Answer: The warranty period for Product A is one year, calculated from the date of purchase." These knowledge items are stored in the professional knowledge base using a specific data structure (such as key-value pairs or JSON format). A professional knowledge base is a database that centrally stores and manages enterprise knowledge and supports rapid retrieval and updating. When storing a knowledge item, the construction agent checks whether the same or similar knowledge item already exists in the knowledge base. If so, the construction agent updates the existing item to ensure the knowledge base is up-to-date. If not, the construction agent creates a new knowledge item. For example, if the knowledge base already has an entry for "Product A's Warranty Period," but the content is "Warranty Period is Six Months," the build agent will update it to "Warranty Period is One Year." In this way, the professional knowledge base can reflect the latest changes in the company's business in real time, providing customer service personnel with accurate and timely knowledge support.

[0080] It should be noted that the construction agent in this embodiment creates a version record for each knowledge entry when storing a new knowledge entry or updating an existing one. For example, when the knowledge entry "Warranty period of Product A" is updated from "six months" to "one year," the construction agent will record two versions: the old version (six months) and the new version (one year). The version management mechanism allows the system to trace back to the old version when needed, such as when the new version information is found to be incorrect. At the same time, version management also supports auditing the change history of knowledge items, making it easier to track the update process of the knowledge base. This mechanism not only improves the reliability and traceability of the knowledge base, but also facilitates the maintenance and management of the knowledge base.

[0081] It should be noted that when the Construction Agent in this embodiment stores knowledge items, it distributes the knowledge base across multiple servers or nodes to improve system availability and fault tolerance. For example, the knowledge base can be deployed in multiple cloud regions, with each region having a knowledge base replica. When the Construction Agent updates a knowledge item, the system automatically synchronizes the update to all replicas, ensuring that the content of each replica is consistent. This distributed storage and synchronization mechanism not only improves the reliability of the knowledge base but also supports large-scale concurrent access, making it particularly suitable for customer service systems of large enterprises or multinational corporations.

[0082] S1600: Utilize the response agent to respond to the actual user business question received and provide a corresponding answer to the customer service based on the professional knowledge base.

[0083] After the construction agent stores similar question-answer pairs summarized based on the structured knowledge items in the professional knowledge base to update the professional knowledge base, the response agent, in response to actual user business questions, provides customer service representatives with corresponding answers based on the professional knowledge base. The specific process is as follows: When the customer service system receives a new question from a user, the response agent first retrieves the knowledge item from the professional knowledge base that best matches the question. For example, a user asks, "What is the warranty period for Product A?" The response agent uses natural language processing technology to analyze the semantics of the question, extract key information (such as "Product A" and "warranty period"), and then searches the knowledge base for the corresponding item. Suppose the knowledge base contains an item, "Question: What is the warranty period for Product A? Answer: The warranty period for Product A is one year, calculated from the date of purchase." The response agent extracts the answer to this item and provides it to the customer service representative in a user-friendly format. The customer service representative can directly use this answer to respond to the user or fine-tune it based on the actual situation. In addition, the response agent records the retrieval and usage of the question, allowing for subsequent analysis of the frequency and effectiveness of knowledge base usage. In this way, the response agent can quickly and accurately provide support to customer service representatives, improving customer service efficiency and user experience.

[0084] Furthermore, when searching the knowledge base, the response agent not only provides the most relevant answer but also recommends multiple relevant answers for customer service representatives to choose from, based on the semantics and context of the question. For example, for a user question like "What is the warranty period for Product A?", the response agent not only provides the directly matching answer but also recommends answers to other frequently asked questions related to "Product A," such as "Product A's return and exchange policy" or "Product A's repair process." Furthermore, the response agent dynamically adjusts its recommendation strategy based on the customer service representative's choices and usage, prioritizing answers that are frequently used and have positive user feedback. This intelligent recommendation mechanism helps customer service representatives more comprehensively understand user needs and provide higher-quality customer service.

[0085] In a specific implementation scenario, an e-commerce company's customer service system handles a large number of user inquiries daily, covering a wide range of issues, including product information, order status, and after-sales service. To improve customer service efficiency and quality, the company has introduced a customer service question-and-answer processing system based on a multi-agent large-scale model. This system comprises a recognition agent, an evaluation agent, a construction agent, and a response agent, responsible for collecting and preprocessing conversation data, evaluating question-and-answer pairs, building and updating the knowledge base, and providing customer service support. For example, when a user initiates a consultation through the online customer service system, the recognition agent captures the conversation data in real time. For example, a user might ask, "It's been a week since I purchased product A, and I haven't received it yet. What should I do?" The recognition agent preprocesses the conversation text, removing irrelevant symbols and breaking it into sentences, generating clear conversation text. The recognition agent uses a large-scale model fine-tuned for the customer service domain to perform semantic analysis on the preprocessed text, identifying the core content of the user's question and labeling it as an "order delay" issue. The evaluation agent evaluates the question-and-answer pairs based on pre-set criteria. If the customer service replies, "Your order is in transit and expected to arrive within 3 days," the judging agent will, based on the accuracy of the answer and user satisfaction, mark the question-and-answer pair as "Perfectly Solved." Construction Agent: The construction agent will categorize and organize the question-and-answer pairs marked as perfectly solved, organizing them into structured knowledge items, such as "Question: What if my order is delayed? Answer: Your order is in transit and expected to arrive within 3 days," and store them in the knowledge base. Response Agent: When the customer service replies to a similar question, the response agent will retrieve the item from the knowledge base and provide a reference answer, helping the customer service replies quickly and improving service efficiency. This multi-agent collaborative customer service Q&A processing approach allows e-commerce companies to significantly improve customer service efficiency, reduce user wait times, increase user satisfaction, and optimize the overall customer service experience.

[0086] In the above implementation, based on the collaborative work of multiple agents and the use of natural language processing technology based on large models, user questions are accurately identified and understood. At the same time, the question and answer evaluation agent comprehensively evaluates the customer service answers, ensuring the high quality and consistency of the answers, and making up for the defect of uneven answer quality in traditional customer service systems. In addition, the knowledge base is used to build an agent that can integrate high-quality question and answer pairs in real time and dynamically update the knowledge base to make its content more timely and practical, solving the problems of slow updates and outdated content in the existing knowledge base, improving the efficiency and accuracy of customer service answers to user questions, and significantly improving the intelligence level of the customer service system and user experience.

[0087] In some implementations, S1200 uses the recognition agent to perform semantic analysis on the pre-processed conversation text, identify the user question and the customer service response, and generate a question-answer pair containing the problem resolution status, including:

[0088] S1211. Using a large model fine-tuned in the customer service domain, perform intent recognition and semantic segmentation on the pre-processed conversation text to distinguish the boundary between user questions and customer service answers;

[0089] In this embodiment, the recognition agent performs semantic analysis on the preprocessed conversation text, identifies the user question and the customer service response, and generates a question-answer pair containing the problem resolution status. First, the recognition agent uses a large model fine-tuned for the customer service domain to perform intent recognition and semantic segmentation on the preprocessed conversation text, distinguishing the boundary between the user question and the customer service response. Specifically, after receiving the preprocessed conversation text, the recognition agent invokes the large model fine-tuned for the customer service domain to perform in-depth analysis of the text. Using natural language processing technology, the large model can identify the intent in the text and perform semantic segmentation on the text. For example, for the preprocessed text "I would like to inquire about how to use Product A. For instructions on how to use Product A, please refer to page 10 of the user manual." The large model first identifies the user's intent as "inquiring about how to use Product A." It then performs semantic segmentation on the text, extracting the user's question and the customer service response, respectively. The user's question is labeled "How to use Product A," and the customer service response is labeled "For instructions on how to use Product A, please refer to page 10 of the user manual." In this way, the recognition agent can clearly distinguish the boundary between the user's question and the customer service response, providing accurate input data for subsequent semantic matching and problem resolution status determination.

[0090] S1212. Match the user question with the customer service answer based on context association analysis;

[0091] After using a large model fine-tuned for the customer service domain to perform intent recognition and semantic segmentation on the pre-processed conversation text and distinguish the boundaries between the user question and the customer service response, this embodiment also matches the user question with the customer service response based on contextual association analysis. Specifically, after completing intent recognition and semantic segmentation, the recognition agent extracts key information from the user question and the customer service response. Then, using contextual association analysis technology, combined with the contextual information of the conversation, the user question and the customer service response are matched. For example, if the user question is "How do I use Product A?" and the customer service response is "For instructions on how to use Product A, please refer to page 10 of the user manual," the recognition agent will analyze keywords (such as "Product A" and "How to use") in the question and answer, as well as contextual information (such as the chronological order of the conversation and the context of the conversation) to confirm that the two sentences are semantically related. In this way, the recognition agent can accurately match the user question with the customer service response, generating a complete question-answer pair. This matching method based on contextual association analysis can effectively handle complex conversation situations, such as multiple rounds of dialogue and follow-up questions, ensuring the accuracy and completeness of the question-answer pair.

[0092] Furthermore, after extracting key information from the user's question and the customer service response, the recognition agent uses natural language processing techniques to calculate the semantic similarity between them. Specifically, the recognition agent converts the user's question and the customer service response into word vectors and then calculates the cosine similarity between them. For example, for the user question "How do I use Product A?" and the customer service response "For instructions on how to use Product A, please refer to page 10 of the user manual," the calculated cosine similarity is high, indicating that the response and question are highly semantically related. This method allows the recognition agent to more accurately match user questions and customer service responses, especially when the question and response are not exactly identical. This method improves the accuracy and robustness of matching.

[0093] S1213. Determine the problem resolution status through a logical reasoning model. If the customer service answer covers the requirements of the user's problem and there is no contradiction, mark it as resolved and generate a question-answer pair containing the problem resolution status.

[0094] After matching the user question with the customer service response based on contextual analysis, the problem resolution status is determined using a logical reasoning model. If the response covers the user's question and there are no contradictions, the response is marked as resolved, generating a question-answer pair containing the problem resolution status. Specifically, after matching the user question with the customer service response, the recognition agent uses a logical reasoning model to analyze the question-answer pair. Based on pre-set rules and semantic understanding capabilities, the logical reasoning model determines whether the customer service response covers the user's question and there are no contradictions. For example, if the user's question is "What is the warranty period for Product A?" and the customer service response is "The warranty period for Product A is one year, starting from the date of purchase," the logical reasoning model analyzes whether the response accurately and completely addresses the user's question. If the response meets the criteria, the question-answer pair is marked as "resolved"; otherwise, it is marked as "unresolved." Finally, the recognition agent generates a question-answer pair containing the problem resolution status and passes it to subsequent modules. This logical reasoning model effectively evaluates the quality of question-answer pairs, ensuring the high quality of the pairs stored in the knowledge base, thereby improving the overall system performance and user experience.

[0095] To further optimize the results, the system in this embodiment also incorporates user feedback as a basis for judgment. After generating a question-and-answer pair, the recognition agent collects user feedback on the user's satisfaction with the answer. For example, after the conversation, the user can rate the customer service representative's answer (e.g., "satisfied" or "unsatisfied"). The logical reasoning model combines user feedback to conduct a comprehensive evaluation of the question-and-answer pair. If the user is satisfied with the answer, even if the answer is slightly semantically deficient, the model will tend to mark it as "resolved." This approach can more realistically reflect user needs and experiences, further improving the accuracy and reliability of question-and-answer quality assessments.

[0096] To further optimize the results, in one embodiment, the Agent analyzes multiple dimensions when evaluating question-answer pairs, such as accuracy, completeness, timeliness, and user satisfaction, and assigns different weights to each dimension. For example, accuracy may be given a higher weight, while timeliness may be given a lower weight. The logical reasoning model determines whether the question-answer pair has reached the "resolved" status based on the weighted comprehensive score. This approach can more comprehensively evaluate the quality of question-answer pairs, ensuring that the generated question-answer pairs meet high-quality standards in multiple aspects, thereby improving the overall performance of the system and user experience.

[0097] In some embodiments, the preset evaluation criteria include at least one of the following dimensions:

[0098] Accuracy evaluation: Use a large model to verify the consistency of the customer service answer with the domain knowledge base;

[0099] Completeness evaluation: Analyze whether the customer service response covers all sub-requirements of the user's question;

[0100] Timeliness evaluation: Compare customer service response time with preset thresholds;

[0101] User satisfaction prediction: Outputs satisfaction scores based on the conversation sentiment analysis model.

[0102] In this embodiment, the judging agent performs a multi-dimensional evaluation of the question-answer pair based on preset judging criteria to determine whether the customer service answer perfectly solves the user's problem, referring to at least one of the following dimensions:

[0103] (1) Accuracy evaluation. In this embodiment, the accuracy evaluation verifies the consistency between the customer service answer and the domain knowledge base through the big model. The specific process is as follows: After the evaluation agent receives the question and answer pair, it calls the big model fine-tuned based on the customer service domain and compares the customer service answer with the standard answer in the knowledge base. For example, the user question is "How long is the warranty period for product A?" The standard answer in the knowledge base is "The warranty period for product A is one year, calculated from the date of purchase." The customer service answer is "The warranty period for product A is one year." The evaluation agent uses the big model to analyze the semantic consistency between the two. Through keyword matching and semantic understanding, it confirms that the customer service answer is highly consistent with the knowledge base answer, and thus determines that the answer meets the standard in terms of accuracy. This method can ensure that the customer service answer is based on accurate knowledge and improve the credibility and authority of the answer.

[0104] (2) Completeness evaluation. In this embodiment, the completeness evaluation analyzes whether the customer service answer covers all the sub-requirements of the user's question. The specific process is as follows: After the evaluation agent receives the question-answer pair, it uses natural language processing technology to break down the user's question into sub-requirements. For example, the user question is "What are the usage and precautions of product A?" The evaluation agent identifies two sub-requirements: "usage method" and "precautions." Then, the evaluation agent analyzes whether the customer service answer covers both sub-requirements. If the customer service answer only mentions "usage method" but not "precautions", it is determined that the answer does not meet the standard in the completeness dimension. In this way, the completeness evaluation can ensure that the customer service answer comprehensively solves the user's problem and improves the user experience.

[0105] (3) Timeliness evaluation. In this embodiment, the timeliness evaluation statistically compares the customer service response time with the preset threshold. The specific process is as follows: After the evaluation agent receives the question-answer pair, it records the time difference from the customer service receiving the user's question to the time the customer service gives the answer. For example, the preset response time threshold is 3 minutes, and the actual response time is 2 minutes. The evaluation agent compares the actual response time with the threshold. If the actual time is less than or equal to the threshold, the answer is judged to meet the standard in terms of timeliness; otherwise, it is judged to be unsatisfactory. This method can ensure that customer service responds to user needs within a reasonable time, improving service efficiency and user satisfaction.

[0106] (4) User satisfaction prediction. In this embodiment, user satisfaction prediction is based on the output of a satisfaction score based on the conversation sentiment analysis model. The specific process is as follows: After the judging agent receives the question-answer pair, it calls the sentiment analysis model to analyze the entire conversation text. For example, if the user says in the conversation, "Thank you very much, this answer is very clear!" the sentiment analysis model will identify the user's emotional tendency as "satisfied" and output a high satisfaction score. The judging agent judges the user's satisfaction with the customer service answer based on the score. This method can evaluate the quality of the answer from the user's perspective and provide a basis for the continuous optimization of customer service quality.

[0107] This embodiment comprehensively and objectively evaluates customer service quality by comprehensively evaluating customer service responses based on multiple dimensions, including accuracy, completeness, timeliness, and user satisfaction prediction. This allows the system to accurately identify high-quality customer service responses and incorporate them into the knowledge base, providing a reference for subsequent customer service work. This not only improves the practicality of the knowledge base but also optimizes the user experience, increases customer service efficiency and satisfaction, and significantly enhances the system's service quality and intelligence.

[0108] In some implementations, S1400 utilizes a construction agent to classify and organize the set of question-answer pairs marked as perfectly solved, so as to summarize similar question-answer pairs and extract structured knowledge items, including:

[0109] S1411. Perform semantic vectorization on the question-answer pairs marked as perfectly solved, and use a clustering algorithm to divide similar question-answer pairs.

[0110] In this embodiment, the construction agent classifies and organizes the set of question-answer pairs marked as perfectly solved to summarize similar question-answer pairs and extract structured knowledge items. This process also includes semantic vectorization of the question-answer pairs marked as perfectly solved, and the use of a clustering algorithm to group similar question-answer pairs. Specifically, after receiving the set of question-answer pairs marked as perfectly solved, the construction agent first uses natural language processing technology to convert the text content of each question-answer pair into a semantic vector. For example, for the question-answer pair "Question: What is the warranty period for Product A? Answer: The warranty period for Product A is one year, calculated from the date of purchase," the construction agent encodes the text content into a high-dimensional semantic vector using a pre-trained large model (such as BERT or GPT). These semantic vectors are then clustered using a clustering algorithm (such as K-Means or DBSCAN) to group semantically similar question-answer pairs into the same category. For example, all question-answer pairs regarding "Product A's warranty period" are clustered into a single group. In this way, the constructed agent can efficiently identify and classify similar question-answer pairs, laying the foundation for further knowledge extraction and structured processing.

[0111] In this embodiment, the semantic vectorization process incorporates multimodal feature fusion. The agent construction not only considers the text content but also incorporates other modal information related to the question-answer pair, such as voice intonation, user expressions (if the conversation includes video or image information), and so on. For example, in a voice customer service conversation, the agent construction extracts features of the voice signal (such as intonation and speaking rate) and fuses them with the semantic vector of the text content. A clustering algorithm is then used to perform cluster analysis on the fused multimodal feature vector. This approach can more comprehensively capture the semantic information of the question-answer pair, improving the accuracy and robustness of clustering, and is particularly suitable for question-answer pair classification in complex scenarios.

[0112] Furthermore, when performing cluster analysis, the Construction Agent dynamically adjusts clustering algorithm parameters, such as the number of cluster centers (K value) or the density threshold, based on the characteristics of the question-answer pair set. For example, if a large number of semantically similar question-answer pairs exist, the Construction Agent automatically increases the number of cluster centers to achieve a finer classification. Conversely, if the semantic diversity of the question-answer pairs is high, the Construction Agent reduces the number of cluster centers to avoid over-segmentation. By dynamically adjusting parameters, the Construction Agent can better adapt to question-answer pairs of varying sizes and complexities, improving clustering efficiency and classification quality.

[0113] S1412. Based on the text summarization capability of the large model, extract the common answer framework and differences from each group of similar question and answer pairs;

[0114] After semantic vectorization of question-and-answer pairs marked as perfectly solved and clustering algorithms to identify similar pairs, the large-scale model's text summarization capabilities are leveraged to extract common answer frameworks and differences from each set of similar pairs. Specifically, after completing cluster analysis, the agent uses a large-scale model fine-tuned for the customer service domain (such as GPT or BERT) to perform text summarization on the answers for each set of similar pairs. For example, for a question-and-answer pair regarding "Product A's warranty period," the large-scale model analyzes multiple customer service answers, extracting common content (such as "Product A's warranty period is one year, starting from the date of purchase") and identifying differences, such as the mention of "For specific terms, please refer to page 10 of the user manual" in some answers. This approach allows the agent to generate a common answer framework while retaining key differences. The common answer framework is "Product A's warranty period is one year, starting from the date of purchase," and the difference is "For specific terms, please refer to page 10 of the user manual." This processing approach effectively extracts structured knowledge items, providing high-quality content for knowledge base updates.

[0115] It's worth noting that when analyzing each set of similar question-answer pairs, the construction agent draws on a library of predefined answer templates. These templates contain standard answer frameworks for common questions. For example, for the question "Product A's after-sales service," the template library may already have a standard answer framework defined: "Product A's after-sales service includes [service content]. For specific terms, please refer to [document location]." The construction agent matches each customer service answer against the template, extracting matching parts as the common answer framework and recording mismatching parts as differences. In this way, the construction agent can quickly generate structured knowledge entries while ensuring the standardization and consistency of the knowledge entries.

[0116] S1413: Merge the common answer framework and the difference points to generate structured knowledge items covering multiple scenarios;

[0117] In this embodiment, based on the text summarization capabilities of a large model, after extracting common answer frameworks and differences from each set of similar question-answer pairs, the common answer frameworks and differences are merged to generate structured knowledge items covering multiple scenarios. Specifically, after extracting the common answer frameworks and differences, the construction agent integrates them into a complete knowledge item. For example, for the question "Warranty period of Product A," the common answer framework is "The warranty period of Product A is one year, calculated from the date of purchase," and the difference is "For specific terms, please refer to page 10 of the user manual." The construction agent merges these two parts to generate a structured knowledge item: "Question: What is the warranty period of Product A? Answer: The warranty period of Product A is one year, calculated from the date of purchase. For specific terms, please refer to page 10 of the user manual." This structured knowledge item not only covers the core information but also provides additional details to meet user needs in different scenarios. Ultimately, these knowledge items are stored in a professional knowledge base, providing customer service personnel with a comprehensive and accurate reference.

[0118] It should be noted that when building an agent, when merging the common answer framework and the difference points, multiple sub-entries will be generated based on the nature of the difference points to cover different scenarios. For example, for the question "How to use Product A," the common answer framework is "The basic usage methods of Product A include powering on, setting up, and operating," and the difference point is "The specific setting methods are different for different models." Building an agent will generate multiple sub-entries, providing detailed usage methods for different models of Product A. For example: "Question: How to use Product A? Answer: The basic usage methods of Product A include powering on, setting up, and operating. For specific setting methods, please refer to the following model instructions: Setting methods for Model A... Setting methods for Model B..." This method can ensure the applicability of knowledge items in different scenarios and improve the practicality and flexibility of the knowledge base.

[0119] In some implementations, after S1413 combines the common answer framework with the difference points to generate structured knowledge items covering multiple scenarios, the process further includes:

[0120] S1421: Modeling the dependency relationships between the structured knowledge items through a graph neural network (GNN) to generate a knowledge graph;

[0121] In this embodiment, after merging the common answer framework with the differences to generate structured knowledge items covering multiple scenarios, the process also includes modeling the dependencies between the structured knowledge items using a graph neural network (GNN) to create a knowledge graph. Specifically, after generating the structured knowledge items covering multiple scenarios, the construction agent uses a graph neural network (GNN) to model the semantic associations and logical dependencies between these knowledge items. For example, for the knowledge items "Warranty period of product A" and "Return and exchange policy of product A," the GNN identifies the association between them (e.g., both fall under the after-sales service category) and connects them as nodes. Each knowledge item is a node, and the edges between nodes represent the strength of their semantic associations. In this way, the construction agent generates a knowledge graph, where nodes represent knowledge items and edges represent dependencies between items. The knowledge graph can intuitively display the associations between knowledge items, facilitate subsequent retrieval and recommendation, and improve the efficiency of knowledge base organization and management.

[0122] Furthermore, after generating the knowledge graph, the constructed agent monitors updates to the knowledge base and changes in user queries in real time. For example, when a new knowledge entry about "Software Updates for Product A" is added to the knowledge base, the dynamic GNN automatically identifies its connections to existing knowledge entries (such as "How to Use Product A") and integrates them into the knowledge graph. Furthermore, if user query frequency changes, the dynamic GNN adjusts the weights of nodes in the knowledge graph in real time to optimize search priority. This approach ensures that the knowledge graph is always up-to-date, improving the system's responsiveness and adaptability.

[0123] S1422: Dynamically adjust the node weights of the knowledge graph according to the user query frequency to optimize the retrieval priority.

[0124] In this embodiment, after generating a knowledge graph by modeling the dependencies between structured knowledge items using a graph neural network (GNN), the node weights of the knowledge graph are dynamically adjusted based on user query frequency to optimize search priority. Specifically, after generating the knowledge graph, the construction agent monitors the frequency of user queries for knowledge items in real time. For example, if the knowledge item "Product A's warranty period" is frequently queried, the construction agent will increase the weight of this node in the knowledge graph. By analyzing query logs, the system finds that "How to use Product A" and "Product A's warranty period" are the two most frequently queried questions by users, so the weights of these two nodes are significantly increased. In the knowledge graph, nodes with higher weights are prioritized for customer service personnel or users, thereby optimizing search efficiency. For example, when a user enters a question related to "Product A," the system will prioritize recommended knowledge items with higher weights, ensuring that the user can quickly access the most relevant information. This approach dynamically adjusts the knowledge graph based on actual user needs, improving the system's intelligence and user experience.

[0125] Furthermore, when building an agent to monitor user query frequency, the system considers the impact of time on weighting. For example, knowledge items that have been frequently queried within the past week will receive a higher weight, while knowledge items that have not been queried for a long time will have their weights gradually reduced. This time-decay model ensures that node weights in the knowledge graph reflect the latest user needs and avoids weight imbalances caused by the accumulation of historical data. For example, if a new feature is released and user query frequency for it increases significantly, the system will promptly adjust the weights to prioritize content related to the new feature.

[0126] In some implementations, S1500 utilizes the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base, and further includes:

[0127] S1511: Adding a version identifier and an effective timestamp to the newly generated structured knowledge entry;

[0128] S1512: Regularly detect the reference frequency of each structured knowledge item in the professional knowledge base according to the version identifier and the effective timestamp, and archive items with a reference frequency lower than a preset value into a low-frequency item history library;

[0129] S1513: Update the priority retrieval rule based on the reference frequency.

[0130] In this embodiment, the construction agent is used to store the summarized similar question-answer pairs based on the structured knowledge items in the professional knowledge base to update the professional knowledge base. This also includes adding a version identifier and an effective timestamp to the newly generated structured knowledge items. Specifically, before storing the structured knowledge items in the professional knowledge base, the construction agent first generates a unique version identifier and effective timestamp for each knowledge item. For example, for the newly generated knowledge item "Question: What is the warranty period for Product A? Answer: The warranty period for Product A is one year, calculated from the date of purchase," the construction agent will assign it a version identifier "V1.0" and an effective timestamp "2024-10-01 10:00:00." The version identifier is used to distinguish different versions of the same knowledge item, while the effective timestamp records the specific time when the version took effect. In this way, the construction agent can ensure that each knowledge item in the knowledge base has a clear version and effective time record, facilitating subsequent management and updating. Furthermore, the version identifier and effective timestamp provide a basis for auditing and tracing knowledge items.

[0131] Furthermore, after adding a version identifier and effective timestamp to the newly generated structured knowledge item, the construction agent regularly checks the reference frequency of each structured knowledge item in the professional knowledge base based on the version identifier and effective timestamp. Items with a reference frequency below a preset value are archived to a low-frequency item history library. Specifically, after storing the structured knowledge item in the professional knowledge base, the construction agent periodically (e.g., weekly or monthly) checks the reference frequency of the items in the knowledge base. Reference frequency refers to the number of times each knowledge item is queried or referenced within a certain period of time. For example, for the knowledge item "Question: What is the warranty period for Product A? Answer: The warranty period for Product A is one year, calculated from the date of purchase," the construction agent counts its references over the past month. If the reference count falls below a preset threshold (e.g., fewer than 10 references per month), the knowledge item is marked as a low-frequency item and archived to a history library. The history library is used to store infrequently used knowledge items for backtracking and auditing when needed, and also provides data support for knowledge base optimization. In this way, the agent can dynamically manage the content of the knowledge base, ensuring that the knowledge base retains the knowledge items that are frequently used and most valuable to users, thereby improving the efficiency and practicality of the knowledge base.

[0132] Furthermore, the citation frequency of each structured knowledge item in the professional knowledge base is regularly monitored based on the version identifier and effective timestamp. After archiving items with citation frequencies below a preset value into a low-frequency item history library, priority search rules are then updated based on the citation frequencies. Specifically, after completing the citation frequency monitoring for each structured knowledge item in the professional knowledge base, the agent dynamically adjusts the search priority rules based on the citation frequencies. For example, for knowledge items with high citation frequencies, such as "Question: How do I use Product A?" (cited over 100 times per month), the system will raise their retrieval priority to the highest level. For knowledge items with low citation frequencies, such as "Question: What are the advanced feature settings for Product A?" (cited less than 10 times per month), the system will lower their retrieval priority. When users or customer service personnel initiate queries, the system will prioritize and recommend high-priority knowledge items, thereby improving retrieval efficiency and user experience. For example, when a user enters a question related to "Product A," the system will prioritize frequently used knowledge items with high user satisfaction, ensuring that users can quickly obtain the most relevant information.

[0133] In this way, the constructed Agent can dynamically optimize the retrieval rules of the knowledge base according to actual usage, thereby improving the intelligence level and response speed of the system.

[0134] In some implementations, S1100 pre-processes the conversation data, including:

[0135] S1121, removing noise from the conversation data, deleting irrelevant symbols and non-text information, and obtaining a conversation text;

[0136] S1122, dividing the conversation text into sentences according to semantic coherence, and converting the sentences into a unified coding format;

[0137] S1123. Based on the conversation text in the unified coding format, extract the scene tags of the conversation text, and associate the scene tags with the conversation time and user identity information to generate a preprocessed conversation text.

[0138] In this embodiment, the conversation data is preprocessed, including noise removal and deletion of irrelevant symbols and non-text information, to obtain the conversation text. Specifically, after receiving the raw conversation data, the recognition agent first uses a text cleaning algorithm to identify and remove irrelevant symbols (such as unnecessary punctuation, special characters, etc.) and non-text information (such as emoticons, image links, etc.). For example, for the original conversation "Hello! [emoticon] I would like to inquire about the usage of product A. Thank you. [image link]", the recognition agent would remove the emoticon "emoticon" and the image link, retaining only the text content "Hello! I would like to inquire about the usage of product A. Thank you." In this way, the recognition agent can remove noise from the conversation data and extract clean, clear conversation text, providing high-quality input data for subsequent semantic analysis and processing.

[0139] Furthermore, the conversation text is segmented into sentences based on semantic coherence and converted into a unified encoding format. Specifically, after noise removal, the recognition agent uses natural language processing techniques to segment the conversation text. For example, for the text "Hello! I would like to inquire about the usage of Product A, thank you.", the recognition agent will segment it into three separate sentences: "Hello!", "I would like to inquire about the usage of Product A," and "Thank you." The recognition agent then converts these sentences into a unified encoding format (such as UTF-8) to ensure that the text can be correctly recognized and processed in subsequent processing. In this way, the recognition agent can standardize the conversation text, providing standardized input data for subsequent semantic analysis and question-answer pair generation.

[0140] Finally, based on the conversation text in the unified encoding format, the recognition agent extracts the scene label from the conversation text and associates it with the conversation time and user identity information to generate preprocessed conversation text. Specifically, after completing sentence segmentation and encoding conversion of the conversation text, the recognition agent uses a pretrained classification model to extract the scene label from the conversation text. For example, for the text "I would like to consult about the usage of Product A," the recognition agent extracts the scene label "Product Consultation." Simultaneously, the recognition agent records the time of the conversation (e.g., "2024-10-01 10:00:00") and user identity information (e.g., user ID "12345"). The recognition agent then associates the scene label, conversation time, and user identity information with the conversation text to generate preprocessed conversation text. For example, the final preprocessing result is: "User ID: 12345, Conversation Time: 2024-10-01 10:00:00, Scene Label: Product Consultation, Conversation Text: I would like to consult about the usage of Product A."

[0141] Through the above implementation, the recognition agent can provide richer context information for subsequent semantic analysis and knowledge base construction, thereby improving the intelligence level of the system and user experience.

[0142] In some implementations, due to the wide variety of user types, different users may have different responses and needs for the same question. Using a standardized answer may produce different results for different users. Therefore, balancing standardized results with user personalization is a challenging issue in the customer service Q&A field. Steps S1600 and later include:

[0143] S1711. Obtaining actual user feedback and user portraits;

[0144] After providing the customer service with the corresponding answers through the professional knowledge base, feedback information from actual users is collected. The feedback information includes: the actual users' evaluation of the answers, as well as the objections raised by the actual users or areas for improvement.

[0145] While obtaining actual user feedback, a user profile can also be collected. This can be done by using existing user profiles in the user's historical data or by generating a user profile based on the user's behavior and emotional changes during the Q&A process. In some embodiments, to quickly extract a profile for a real user, a historical user with a complete user profile is used as a standard profile. Specifically, based on the questions a real user asked a customer service representative, a user profile of a historical user with the same question and complete profile data is extracted from the historical data as the standard profile. Based on the differences in voice, body language, and facial expressions between the real user and the customer service representative during their interaction, the standard profile is corrected for the differences between the two users for the same question, resulting in a user profile for the real user. Because some users may have zero interaction with customer service representatives, randomly or ad hoc user profiles are often inaccurate or lack sufficient information. Using this method, a user profile for a real user can be quickly generated. Furthermore, the differential information between the two users for the same question best reflects the differences between the users. Therefore, the accuracy of the user profiles obtained using this method is greatly improved.

[0146] S1712: Perform personalized adjustments on the target structured knowledge items in the professional knowledge base according to the feedback information to generate derived structured knowledge items of the target structured knowledge items;

[0147] The target structured knowledge item in this embodiment refers to a structured knowledge item corresponding to a specific problem of a real user.

[0148] Based on the user's dissatisfaction with the target structured knowledge item or modification suggestions in the feedback information, the target structured knowledge item is corrected and fine-tuned to generate personalized derived structured knowledge items for real users that can meet the specific requirements of real users.

[0149] S1713: placing the user portrait in a preset multi-dimensional portrait coordinate system, and obtaining the portrait coordinates of the user portrait;

[0150] After obtaining the user profile, place it in a multidimensional profile coordinate system. This pre-built system is used to store the spatial coordinates of multiple user profiles. Each axis in the system represents a dimension of the user profile, such as age, personality, and emotions. Depending on the actual dimensionality requirements, the system can be two-dimensional, three-dimensional, or four-dimensional.

[0151] The performance of each dimension in the user portrait can be quantified as the coordinate value on that dimension. According to the coordinate value of each dimension in the user portrait, the portrait coordinates can be obtained, and the user portrait can be projected into a multi-dimensional portrait coordinate system based on the user portrait coordinates.

[0152] S1714. Associate the derived structured knowledge item with the portrait coordinates to generate a multi-dimensional knowledge base with the target structured knowledge item as the coordinate zero point, the portrait coordinates as variables, and the derived structured knowledge item as the distribution point.

[0153] After obtaining the portrait coordinates, the derived structured knowledge items are associated with the portrait coordinates to form associated data. A multidimensional knowledge coordinate system is then constructed based on the portrait coordinates. This multidimensional knowledge coordinate system uses the target structured knowledge item as the coordinate zero point, the portrait coordinates as the dimensional variables, and the derived structured knowledge items as the distribution points, generating a multidimensional knowledge coordinate system with the target structured knowledge item as the origin coordinate. In this multidimensional knowledge coordinate system, the target structured knowledge item and the numerous derived structured knowledge items form a trunk-branch structure similar to a "virus" structure, and this data structure is stored in the multidimensional knowledge base.

[0154] When a new user asks the same question and interacts with the customer service staff, the corresponding portrait coordinates are obtained in the multi-dimensional portrait coordinate system based on the user's user portrait. Then, based on the portrait coordinates, the corresponding derived structured knowledge items are extracted in the multi-dimensional knowledge coordinate system and sent to the customer service staff as the standard answer prompt for the user.

[0155] Traditional customer service Q&A often uses standardized responses, making it difficult to adapt to the diverse responses and needs of different users to the same question. This solution leverages user feedback and profiles to adjust the target structured knowledge items in the professional knowledge base based on the user's individual needs, generating adapted derived structured knowledge items. This allows customer service representatives to provide personalized responses tailored to the user's specific situation, improving user satisfaction with customer service. Based on the user's profile coordinates, customer service representatives can quickly extract the corresponding derived structured knowledge items from the knowledge base and provide accurate responses, improving customer service efficiency and quality while reducing user wait times.

[0156] Please refer to the following for details: Figure 2 , Figure 2 Schematic diagram of the basic structure of the customer service question and answer processing device based on the multi-agent large model in this embodiment.

[0157] like Figure 2 As shown, a customer service question and answer processing device based on a multi-agent large model includes: a first recognition module 1100, which is used to use the recognition agent to obtain the conversation data between the customer service and the user, pre-process the conversation data, and generate a pre-processed conversation text; a second recognition module 1200, which is used to use the recognition agent to perform semantic analysis on the pre-processed conversation text, identify the user question and the customer service answer, and generate a question and answer pair containing the problem solving status; a question and answer evaluation module 1300, which is used to use the evaluation agent to perform a multi-dimensional evaluation on the question and answer pair based on preset evaluation criteria to determine whether the customer service answer is correct. Whether the user problem is perfectly solved, a set of question and answer pairs marked as perfectly solved is generated; a first construction module 1400 is used to use the construction agent to classify and organize the set of question and answer pairs marked as perfectly solved, so as to summarize similar question and answer pairs and refine structured knowledge items; a second construction module 1500 is used to use the construction agent to store the summarized similar question and answer pairs in the professional knowledge base based on the structured knowledge items to update the professional knowledge base; a business response module 1600 is used to use the response agent to respond to the actual user business problem received and provide corresponding answers to the customer service based on the professional knowledge base.

[0158] The above-mentioned customer service question and answer processing device based on a multi-agent large model uses an identification agent to obtain conversation data between customer service and users, pre-processes the conversation data, and generates a pre-processed conversation text; uses the identification agent to perform semantic analysis on the pre-processed conversation text, identifies user questions and customer service answers, and generates question and answer pairs containing the problem resolution status; uses the judgment agent to perform multi-dimensional evaluation on the question and answer pairs based on preset judgment criteria, determines whether the customer service answers perfectly solve the user problems, and generates a set of question and answer pairs marked as perfectly resolved; uses the construction agent to classify and organize the set of question and answer pairs marked as perfectly resolved to summarize similar question and answer pairs and refine structured knowledge items; uses the construction agent to store the summarized similar question and answer pairs in a professional knowledge base based on the structured knowledge items to update the professional knowledge base; uses the response agent to respond to actual user business questions received and provide corresponding answers to customer service based on the professional knowledge base. Based on the collaborative work of multiple agents and the use of natural language processing technology based on large models, user questions can be accurately identified and understood. At the same time, the question and answer evaluation agent conducts a comprehensive evaluation of customer service answers, ensuring the high quality and consistency of the answers, and making up for the uneven answer quality in traditional customer service systems. In addition, the knowledge base is used to build an agent that can integrate high-quality question and answer pairs in real time and dynamically update the knowledge base to make its content more timely and practical, solving the problems of slow updates and outdated content in the existing knowledge base, improving the efficiency and accuracy of customer service answers to user questions, and significantly improving the intelligence level of the customer service system and user experience.

[0159] To solve the above technical problems, the present application also provides a computer device. Figure 3 , Figure 3 This is a basic structural block diagram of the computer device in this embodiment.

[0160] like Figure 3 As shown, a schematic diagram of the internal structure of a computer device. The computer device includes a processor, a non-volatile storage medium, a memory and a network interface connected via a system bus. Among them, the non-volatile storage medium of the computer device stores an operating system, a database and computer-readable instructions, and the database may store a control information sequence. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor may execute a customer service question and answer processing method based on a multi-agent large model. The network interface of the computer device is used to connect and communicate with the terminal. Those skilled in the art will understand that Figure 3The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0161] In this embodiment, the processor is used to execute Figure 2 The memory stores the program code and various data required to execute the specific functions of the first identification module 1100, the second identification module 1200, the question-and-answer evaluation module 1300, the first construction module 1400, the second construction module 1500, and the business response module 1600. The network interface is used to transmit data between user terminals or servers. The memory in this embodiment stores the program code and data required to execute all submodules of the customer service question-and-answer processing device based on the multi-agent large model. The server can call the server's program code and data to execute the functions of all submodules.

[0162] The computer device obtains conversation data between customer service and users by using an identification agent, pre-processes the conversation data, and generates a pre-processed conversation text; uses the identification agent to perform semantic analysis on the pre-processed conversation text, identifies user questions and customer service answers, and generates question-answer pairs containing the problem-solving status; uses the judgment agent to perform multi-dimensional evaluation on the question-answer pairs based on preset judgment criteria, determines whether the customer service answers perfectly solve the user problems, and generates a set of question-answer pairs marked as perfectly solved; uses the construction agent to classify and organize the set of question-answer pairs marked as perfectly solved, so as to summarize similar question-answer pairs and refine structured knowledge items; uses the construction agent to store the summarized similar question-answer pairs in a professional knowledge base based on the structured knowledge items to update the professional knowledge base; uses the response agent to respond to actual user business questions received and provide corresponding answers to customer service based on the professional knowledge base. Based on the collaborative work of multiple agents and the use of natural language processing technology based on large models, user questions can be accurately identified and understood. At the same time, the question and answer evaluation agent conducts a comprehensive evaluation of customer service answers, ensuring the high quality and consistency of the answers, and making up for the uneven answer quality in traditional customer service systems. In addition, the knowledge base is used to build an agent that can integrate high-quality question and answer pairs in real time and dynamically update the knowledge base to make its content more timely and practical, solving the problems of slow updates and outdated content in the existing knowledge base, improving the efficiency and accuracy of customer service answers to user questions, and significantly improving the intelligence level of the customer service system and user experience.

[0163] The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the customer service question and answer processing method based on the multi-agent large model described in any of the above embodiments.

[0164] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0165] Those skilled in the art will appreciate that the steps, measures, and schemes in the various operations, methods, and processes discussed in this application may be interchanged, modified, combined, or deleted. Furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application may also be interchanged, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and schemes in the prior art that are similar to those disclosed in this application may also be interchanged, modified, rearranged, decomposed, combined, or deleted.

[0166] The above description is only part of the implementation methods of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A customer service question and answer processing method based on a multi-agent large model, characterized in that: include: Using the recognition agent to obtain the conversation data between the customer service and the user, preprocessing the conversation data to generate a preprocessed conversation text; The preprocessing process includes removing noise from the conversation data, deleting irrelevant symbols and non-text information, and obtaining the conversation text; The conversation text is divided into sentences according to semantic coherence and converted into a unified coding format; based on the conversation text in the unified coding format, scene tags of the conversation text are extracted, and the scene tags are associated with the conversation time and user identity information to generate a preprocessed conversation text; The preprocessing process further includes: performing emotion tagging on the conversation data to determine the user's emotional state, and performing topic extraction on the conversation data to determine the topic content of the conversation data, and using the emotional state and the topic content as context information of the conversation text; Using the recognition agent to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status; Using the judging agent to perform a multi-dimensional evaluation of the question-answer pair based on preset judging criteria, judging whether the customer service answer perfectly solves the user's problem based on the different weights assigned to each dimension, and generating a set of question-answer pairs marked as perfectly solved based on the context information; The set of question-answer pairs marked as perfectly solved is classified and sorted using a constructed agent to summarize similar question-answer pairs and extract structured knowledge items. This includes: semantic vectorization of multimodal features of the question-answer pairs marked as perfectly solved, and using a clustering algorithm to divide similar question-answer pairs; extracting common answer frameworks and differences from each group of similar question-answer pairs based on the text summarization capability of the large model; and merging the common answer frameworks and differences to generate structured knowledge items covering multiple scenarios. Using the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items covering multiple scenarios into a professional knowledge base, so as to update the professional knowledge base; The response agent is used to respond to the actual user business question received and provide the customer service with a corresponding answer based on the professional knowledge base.

2. The customer service question and answer processing method based on a multi-agent large model according to claim 1 is characterized in that: The recognition agent is used to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status, including: Using a large model fine-tuned in the customer service field, the pre-processed conversation text is subjected to intent recognition and semantic segmentation to distinguish the boundary between user questions and customer service responses. Match the user's question with the customer service answer based on contextual analysis; The problem solving status is determined through a logical reasoning model. If the customer service answer covers the requirements of the user's problem and there is no contradiction, it is marked as solved, and a question-answer pair containing the problem solving status is generated.

3. The customer service question and answer processing method based on a multi-agent large model according to claim 1 is characterized in that: After the response agent receives the actual user business question and provides a corresponding answer to the customer service based on the professional knowledge base, the method includes: Obtain feedback information and user portraits from actual users; Performing personalized adjustments to the target structured knowledge items in the professional knowledge base according to the feedback information to generate derived structured knowledge items of the target structured knowledge items; Placing the user portrait in a preset multi-dimensional portrait coordinate system and obtaining the portrait coordinates of the user portrait; The derived structured knowledge items are associated with the portrait coordinates to generate a multi-dimensional knowledge base with the target structured knowledge item as the coordinate zero point, the portrait coordinates as variables, and the derived structured knowledge items as distribution points.

4. The customer service question and answer processing method based on a multi-agent large model according to claim 1 is characterized in that: After merging the common answer framework with the difference points to generate structured knowledge items covering multiple scenarios, the method further includes: Modeling the dependency relationships between the structured knowledge items through a graph neural network to generate a knowledge graph; The node weights of the knowledge graph are dynamically adjusted according to the user query frequency to optimize the retrieval priority.

5. The customer service question and answer processing method based on a multi-agent large model according to claim 1 is characterized in that: The step of using the construction agent to store the summarized similar question-answer pairs based on the structured knowledge items into a professional knowledge base to update the professional knowledge base further includes: Adding a version identifier and an effective timestamp to the newly generated structured knowledge entry; Regularly detecting the reference frequency of each structured knowledge item in the professional knowledge base according to the version identifier and the effective timestamp, and archiving the items with a reference frequency lower than a preset value into a low-frequency item history library; The priority retrieval rule is updated based on the reference frequency.

6. A customer service question and answer processing device based on a multi-agent large model, characterized in that: include: The first recognition module is used to use the recognition agent to obtain the conversation data between the customer service and the user, pre-process the conversation data, and generate a pre-processed conversation text; The preprocessing process includes removing noise from the conversation data, deleting irrelevant symbols and non-text information, and obtaining the conversation text; The conversation text is divided into sentences according to semantic coherence and converted into a unified coding format; based on the conversation text in the unified coding format, scene tags of the conversation text are extracted, and the scene tags are associated with the conversation time and user identity information to generate a preprocessed conversation text; The preprocessing process further includes: performing emotion tagging on the conversation data to determine the user's emotional state, and performing topic extraction on the conversation data to determine the topic content of the conversation data, and using the emotional state and the topic content as context information of the conversation text; A second recognition module is used to use the recognition agent to perform semantic analysis on the pre-processed conversation text, identify user questions and customer service answers, and generate question-answer pairs containing the problem resolution status; A question-answer evaluation module is used to use the evaluation agent to perform a multi-dimensional evaluation of the question-answer pair based on preset evaluation criteria, determine whether the customer service answer perfectly solves the user's problem based on the different weights assigned to each dimension, and generate a set of question-answer pairs marked as perfectly solved based on the context information; The first construction module is used to classify and organize the set of question-answer pairs marked as perfectly solved using a construction agent to summarize similar question-answer pairs and extract structured knowledge items. This module includes: performing semantic vectorization of multimodal features on the question-answer pairs marked as perfectly solved, and using a clustering algorithm to divide similar question-answer pairs; extracting common answer frameworks and differences from each group of similar question-answer pairs based on the text summarization capability of the large model; and merging the common answer frameworks and differences to generate structured knowledge items covering multiple scenarios. A second construction module is configured to utilize the construction agent to store the summarized similar question-answer pairs into a professional knowledge base based on the structured knowledge items covering multiple scenarios, so as to update the professional knowledge base; The business response module is used to use the response agent to respond to actual user business questions received and provide corresponding answers to customer service based on the professional knowledge base.

7. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the customer service question and answer processing method based on a multi-agent large model as claimed in any one of claims 1 to 5.

8. A storage medium storing computer-readable instructions, characterized in that: When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the customer service question-and-answer processing method based on a multi-agent large model as claimed in any one of claims 1 to 5.

Images