A maintenance consultation intelligent customer service system based on large language model

Abstract

The present invention relates to the field of instant messaging technology, and discloses a maintenance consultation intelligent customer service system based on a large language model, the system is composed of a user input processing and inspection module, an intention recognition and information extraction module, a dynamic knowledge retrieval module, an intelligent reply generation module, and a reply optimization and quality control module. The present invention uses a large language model to perform security and compliance checks on user input, automatically filters inappropriate or potentially harmful content, ensures the security of the conversation, extracts key information, provides accurate information support for subsequent processing, generates targeted preliminary replies based on the conversation status and user input, the system can determine whether the information is complete, and guides the user to supplement information when necessary, improves the effectiveness of the reply, and achieves the beneficial effect of intelligent and accurate human-computer dialogue.

Description

Technical Field

[0001] The present invention relates to the technical field of instant messaging, and in particular to a maintenance consultation intelligent customer service system based on a large language model. Background Art

[0002] With the rapid development of artificial intelligence technology, especially the breakthroughs of large language model technology in the fields of text understanding and knowledge reasoning, intelligent customer service for maintenance consultation has become a key link in the digital transformation of enterprises. Traditional language systems are often unable to accurately understand the complex needs of the public, resulting in customer service answers that are very different from the actual needs of the public. In addition, traditional customer service systems lack sufficient language comprehension capabilities. When faced with personalized questions from users, they can only answer in a fixed way, lacking richness and flexibility, which has a great impact on user experience. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a maintenance consultation intelligent customer service system based on a large language model, which has intelligent and accurate human-computer dialogue, effectively improves the quality of dialogue, enhances user experience, has broad application prospects in the fields of customer service, intelligent assistants, etc., and solves the above-mentioned technical problems.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a maintenance consultation intelligent customer service system based on a large language model, the system consisting of a user input processing and checking module, an intention recognition and information extraction module, a dynamic knowledge retrieval module, an intelligent reply generation module and a reply optimization and quality control module;

[0005] The user input processing and checking module is used to receive user input, perform scenario and compliance checks, analyze the input using a large language model (LLM), filter inappropriate or potentially harmful content, ensure that the content received by the system meets security and compliance standards, and prevent the system from receiving inappropriate content;

[0006] The intention recognition and information extraction module is used to recognize the user's intention and extract key information. It uses the semantic understanding ability of LLM to recognize the user's input and extract information, grasp the user's real needs, and provide information support for subsequent processing;

[0007] The dynamic knowledge retrieval module is used to determine whether external knowledge retrieval is needed based on the intent recognition results and the extracted information, and to retrieve relevant information from the knowledge base using the simcse semantic similarity model and VecDB technology to obtain external knowledge to support the system in generating more accurate answers;

[0008] The intelligent reply generation module is used to generate preliminary replies, generate targeted answers based on the dialogue status, rely on LLM, generate replies based on the current dialogue process, prompt the user to supplement information when the information is incomplete, and provide suggestions or other responses when the information is complete, provide targeted and coherent replies, and improve the effectiveness and fluency of the dialogue;

[0009] The response optimization and quality control module is used to optimize and perform quality control on the generated preliminary responses, automatically correct and optimize the responses, and ensure their accuracy, relevance, and language fluency.

[0010] Preferably, the system further comprises an application method, wherein the application method is:

[0011] Step 1: User input: The system receives the user's initial input;

[0012] Step 2: LLM checks whether the context is relevant, including the first question and topic change: call the large language model API to compare and analyze the user input with the historical context;

[0013] Step 3: If the context is irrelevant: perform global intent recognition and delete irrelevant context

[0014] Step 4: If the context is relevant: update the context and obtain the slot update value;

[0015] Step 5: Check if the slot is complete: Analyze the user input to determine whether the user needs to be redirected;

[0016] Step 6: If the slot is not fully filled: guide the user to re-enter according to the test results. This step is designed with a prompt template in advance;

[0017] Step 7: If the slots are fully filled: call Vector Database to retrieve the corresponding standard fault phenomenon, maintenance spare parts type and other information. The Vector data stored in Vector Datebase is generated by the Bge-large model. At the same time, the self-service maintenance suggestions corresponding to the fault phenomenon are given and pre-stored in VecDB.

[0018] Step 8: If the problem cannot be solved by self-service: prompt the user to enter name, phone number and other information to generate a maintenance work order;

[0019] Step 9: If the problem can be solved automatically: the process ends automatically.

[0020] Preferably, the user input processing and checking module expression is:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] Accepts user input expressions: ;

[0026] Parse the user input expression as: ;

[0027] The filter content expression is: ;

[0028] Make sure the legal expression is: .

[0029] Preferably, in the dynamic knowledge retrieval module:

[0030] ;

[0031] ;

[0032] The similarity expression between key information and each piece of knowledge in the knowledge base is:

[0033]

[0034] in:

[0035] Represents the first Articles of relevant knowledge.

[0036] Preferably, in the dynamic knowledge retrieval module

[0037]

[0038] The formula is used to determine whether the calculated semantic similarity is greater than or equal to the threshold ;

[0039] The expression for retrieving information from a vector database is:

[0040]

[0041] The formula is used to filter out all knowledge from the knowledge base whose semantic similarity with the key information K is higher than the threshold.

[0042] Preferably, the dynamic knowledge retrieval module also includes providing support calculation, and the support calculation expression is:

[0043]

[0044] The formula uses the retrieved relevant knowledge as external knowledge to support the system in generating more accurate answers.

[0045] Preferably, in the intelligent reply generation module:

[0046] : Indicates the current dialogue status; : Indicates that the conversation status and user input The process of inputting into a large language model for analysis and generating responses; : Indicates whether the information is complete: 0 means incomplete, 1 means complete; : Indicates an action that provides suggestions or other responses;

[0047] Generate the initial reply expression:

[0048] .

[0049] Preferably, the expression for checking whether the information is complete in the intelligent reply generation module is:

[0050]

[0051] Indicates checking whether the generated response contains complete information;

[0052]

[0053] If the information is incomplete , the user is prompted to provide additional information;

[0054]

[0055] If the information is complete , then provide suggestions or other responses;

[0056] The final reply expression is generated as:

[0057] .

[0058] Preferably, the optimization expression for the preliminary response in the response optimization and quality control module is:

[0059]

[0060] : represents the initial response generated; : Indicates the final response after optimization and quality control.

[0061] Compared with the prior art, the present invention provides a maintenance consultation intelligent customer service system based on a large language model, which has the following beneficial effects:

[0062] 1. The present invention uses a large language model to perform security and compliance checks on user input through the system, automatically filters inappropriate or potentially harmful content, ensures the security of the conversation, and extracts key information through the system, providing accurate information support for subsequent processing. It uses the SimCSE semantic similarity model and VecDB technology to retrieve knowledge related to user intent and key information from the knowledge base to ensure the accuracy and comprehensiveness of the reply content, and generates targeted preliminary replies based on the conversation status and user input. The system can determine whether the information is complete and guide the user to supplement the information when necessary, thereby improving the effectiveness of the reply and achieving the beneficial effect of intelligent and accurate human-computer dialogue.

[0063] 2. The present invention uses a large language model to intelligently process and security check user input through the system to ensure the appropriateness of the content of the conversation, thereby reducing unnecessary interference and improving the professionalism and quality of the conversation. The system can accurately identify the user's intentions and extract key information, making the response more relevant to user needs and improving the accuracy and effectiveness of the conversation. The system uses advanced knowledge retrieval technology to ensure that the information provided is relevant to user needs, enhancing the comprehensiveness and practicality of the answer. By generating preliminary replies and optimizing them, the system can provide high-quality answers with fluent language and complete information, thereby improving the user's interactive experience. The system can efficiently handle customer inquiries, provide accurate help, reduce the burden of manual customer service, and can handle a large number of customer requests, improving service efficiency and satisfaction. Through intelligent processing and optimization of conversations, the system has broad application prospects in scenarios such as personal assistants and technical support, and can better meet the user's personalized needs and instant service requirements, achieving the beneficial effect of effectively improving the quality of conversations, enhancing user experience, and having broad application prospects in customer service, intelligent assistants and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a schematic diagram of the system of the present invention;

[0065] Figure 2 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] See also Figure 1-Figure 2 , a maintenance consultation intelligent customer service system based on a large language model, the system consists of a user input processing and checking module, an intention recognition and information extraction module, a dynamic knowledge retrieval module, an intelligent reply generation module and a reply optimization and quality control module;

[0068] The user input processing and checking module is used to receive user input, perform scenario and compliance checks, analyze the input using a large language model (LLM), filter inappropriate or potentially harmful content, ensure that the content received by the system meets security and compliance standards, and prevent the system from receiving inappropriate content;

[0069] The intention recognition and information extraction module is used to recognize the user's intention and extract key information. It uses the semantic understanding ability of LLM to recognize the user's input and extract information, grasp the user's real needs, and provide information support for subsequent processing;

[0070] The dynamic knowledge retrieval module is used to determine whether external knowledge retrieval is needed based on the intent recognition results and the extracted information, and to retrieve relevant information from the knowledge base using the simcse semantic similarity model and VecDB technology to obtain external knowledge to support the system in generating more accurate answers;

[0071] The intelligent reply generation module is used to generate preliminary replies, generate targeted answers based on the dialogue status, rely on LLM, generate replies based on the current dialogue process, prompt the user to supplement information when the information is incomplete, and provide suggestions or other responses when the information is complete, provide targeted and coherent replies, and improve the effectiveness and fluency of the dialogue;

[0072] The response optimization and quality control module is used to optimize and perform quality control on the generated preliminary responses, automatically correct and optimize the responses, and ensure their accuracy, relevance, and language fluency.

[0073] Compliance check: The system uses a large language model (LLM) to perform security and compliance checks on user input, automatically filtering inappropriate or potentially harmful content to ensure the safety of conversations.

[0074] Intent Recognition: LLM analyzes user input and accurately identifies the user's intent.

[0075] Information extraction: The system extracts key information and provides accurate information support for subsequent processing.

[0076] Semantic similarity retrieval: Use the SimCSE semantic similarity model and VecDB technology to retrieve knowledge related to user intent and key information from the knowledge base to ensure the accuracy and comprehensiveness of the reply content.

[0077] Initial response generation: Generate targeted initial responses based on the conversation state and user input.

[0078] Information completeness check: The system can determine whether the information is complete and guide the user to supplement the information when necessary to improve the effectiveness of the response.

[0079] Optimize responses: Automatically correct and optimize the generated preliminary responses to ensure the accuracy, relevance, and language fluency of the final responses.

[0080] Preferably, the system further comprises an application method, wherein the application method is:

[0081] Step 1: User input: The system receives the user's initial input;

[0082] Step 2: LLM checks whether the context is relevant, including the first question and topic change: call the large language model API to compare and analyze the user input with the historical context;

[0083] Step 3: If the context is irrelevant: perform global intent recognition and delete irrelevant context

[0084] Step 4: If the context is relevant: update the context and obtain the slot update value;

[0085] Step 5: Check if the slot is complete: Analyze the user input to determine whether the user needs to be redirected;

[0086] Step 6: If the slot is not fully filled: guide the user to re-enter according to the test results. This step is designed with a prompt template in advance;

[0087] Step 7: If the slots are fully filled: call Vector Database to retrieve the corresponding standard fault phenomenon, maintenance spare parts type and other information. The Vector data stored in Vector Datebase is generated by the Bge-large model. At the same time, the self-service maintenance suggestions corresponding to the fault phenomenon are given and pre-stored in VecDB.

[0088] Step 8: If the problem cannot be solved by self-service: prompt the user to enter name, phone number and other information to generate a maintenance work order;

[0089] Step 9: If the problem can be solved automatically: the process ends automatically.

[0090] Preferably, the user input processing and checking module expression is:

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] Accepts user input expressions: ;

[0096] Parse the user input expression as: ;

[0097] The filter content expression is: ;

[0098] Make sure the legal expression is: .

[0099] Preferably, in the dynamic knowledge retrieval module:

[0100] ;

[0101] ;

[0102] The similarity expression between key information and each piece of knowledge in the knowledge base is:

[0103]

[0104] in:

[0105] Represents the first Articles of relevant knowledge.

[0106] Preferably, in the dynamic knowledge retrieval module

[0107]

[0108] The formula is used to determine whether the calculated semantic similarity is greater than or equal to the threshold ;

[0109] The expression for retrieving information from a vector database is:

[0110]

[0111] The formula is used to filter out all knowledge from the knowledge base whose semantic similarity with the key information K is higher than the threshold.

[0112] Preferably, the dynamic knowledge retrieval module also includes providing support calculation, and the support calculation expression is:

[0113]

[0114] The formula uses the retrieved relevant knowledge as external knowledge to support the system in generating more accurate answers.

[0115] Preferably, in the intelligent reply generation module:

[0116] : Indicates the current dialogue status; : Indicates that the conversation status and user input The process of inputting into a large language model for analysis and generating responses; : Indicates whether the information is complete: 0 means incomplete, 1 means complete; : Indicates an action that provides suggestions or other responses;

[0117] Generate the initial reply expression:

[0118] .

[0119] Preferably, the expression for checking whether the information is complete in the intelligent reply generation module is:

[0120]

[0121] Indicates checking whether the generated response contains complete information;

[0122]

[0123] If the information is incomplete , the user is prompted to provide additional information;

[0124]

[0125] If the information is complete , then provide suggestions or other responses;

[0126] The final reply expression is generated as:

[0127] .

[0128] Preferably, the optimization expression for the preliminary response in the response optimization and quality control module is:

[0129]

[0130] : represents the initial response generated; : Indicates the final response after optimization and quality control.

[0131] Intelligent processing and compliance checking: The system uses a large language model to intelligently process and security check user input to ensure the appropriateness of the conversation content, thereby reducing unnecessary interference and improving the professionalism and quality of the conversation.

[0132] Accurate intent recognition and information extraction: The system can accurately identify the user's intent and extract key information, making the response more relevant to the user's needs and improving the accuracy and effectiveness of the conversation.

[0133] Dynamic knowledge retrieval: The system uses advanced knowledge retrieval technologies (such as SimCSE and VecDB) to ensure that the information provided is relevant to user needs, enhancing the comprehensiveness and practicality of the answers.

[0134] Intelligent reply generation and optimization: By generating preliminary replies and optimizing them, the system can provide high-quality answers with fluent language and complete information, thereby improving the user's interactive experience.

[0135] Customer service area: The system can efficiently handle customer inquiries, provide accurate assistance, reduce the burden of manual customer service, and handle a large number of customer requests, thereby improving service efficiency and satisfaction.

[0136] Smart assistant field: Through intelligent processing and optimized conversations, the system has broad application prospects in scenarios such as personal assistants and technical support, and can better meet users' personalized needs and instant service requirements.

[0137] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A maintenance consultation intelligent customer service system based on a large language model, characterized in that: The system consists of a user input processing and checking module, an intention recognition and information extraction module, a dynamic knowledge retrieval module, an intelligent reply generation module, and a reply optimization and quality control module; The user input processing and checking module is used to receive user input, perform scenario and compliance checks, analyze the input using a large language model (LLM), filter inappropriate or potentially harmful content, ensure that the content received by the system meets security and compliance standards, and prevent the system from receiving inappropriate content; The intention recognition and information extraction module is used to recognize the user's intention and extract key information. It uses the semantic understanding ability of LLM to recognize the user's input and extract information, grasp the user's real needs, and provide information support for subsequent processing; The dynamic knowledge retrieval module is used to determine whether external knowledge retrieval is needed based on the intent recognition results and the extracted information, and to retrieve relevant information from the knowledge base using the simcse semantic similarity model and VecDB technology to obtain external knowledge to support the system in generating more accurate answers; The intelligent reply generation module is used to generate preliminary replies, generate targeted answers based on the dialogue status, rely on LLM, generate replies based on the current dialogue process, prompt the user to supplement information when the information is incomplete, and provide suggestions or other responses when the information is complete, provide targeted and coherent replies, and improve the effectiveness and fluency of the dialogue; The response optimization and quality control module is used to optimize and perform quality control on the generated preliminary responses, automatically correct and optimize the responses, and ensure their accuracy, relevance, and language fluency.

2. The maintenance consultation intelligent customer service system based on a large language model according to claim 1, characterized in that: The system also includes an application method, which is: Step 1: User input: The system receives the user's initial input; Step 2: LLM checks whether the context is relevant, including the first question and topic change: call the large language model API to compare and analyze the user input with the historical context; Step 3: If the context is irrelevant: perform global intent recognition and delete irrelevant context Step 4: If the context is relevant: update the context and obtain the slot update value; Step 5: Check if the slot is complete: Analyze the user input to determine whether the user needs to be redirected; Step 6: If the slot is not fully filled: guide the user to re-enter according to the test results. This step is designed with a prompt template in advance; Step 7: If the slots are fully filled: call Vector Database to retrieve the corresponding standard fault phenomenon, maintenance spare parts type and other information. The Vector data stored in Vector Datebase is generated by the Bge-large model. At the same time, the self-service maintenance suggestions corresponding to the fault phenomenon are given and pre-stored in VecDB. Step 8: If the problem cannot be solved by self-service: prompt the user to enter name, phone number and other information to generate a maintenance work order; Step 9: If the problem can be solved automatically: the process ends automatically.

3. The maintenance consultation intelligent customer service system based on a large language model according to claim 1, characterized in that: The user input processing and checking module expression is: ； ； ； ； Accepts user input expressions: ; Parse the user input expression as: ; The filter content expression is: ; Make sure the legal expression is: .

4. The maintenance consultation intelligent customer service system based on a large language model according to claim 1, characterized in that: In the dynamic knowledge retrieval module: ； ； The similarity expression between key information and each piece of knowledge in the knowledge base is: ,in: Represents the first Articles of relevant knowledge.

5. The maintenance consultation intelligent customer service system based on a large language model according to claim 4 is characterized by: The dynamic knowledge retrieval module determines whether the threshold expression is met: , the formula is used to determine whether the calculated semantic similarity is greater than or equal to the threshold ; The expression for retrieving information from a vector database is: The formula is used to filter out all knowledge from the knowledge base whose semantic similarity with the key information K is higher than the threshold.

6. The maintenance consultation intelligent customer service system based on a large language model according to claim 5, characterized in that: The dynamic knowledge retrieval module also includes providing support calculation, and the support calculation expression is: , the formula uses the retrieved relevant knowledge as external knowledge to support the system in generating more accurate answers.

7. The maintenance consultation intelligent customer service system based on a large language model according to claim 3 is characterized by: In the smart reply generation module: : Indicates the current dialogue status; : Indicates that the conversation status and user input The process of inputting into a large language model for analysis and generating responses; : Indicates whether the information is complete: 0 means incomplete, 1 means complete; : Indicates an action that provides suggestions or other responses; Generate the initial reply expression: .

8. The maintenance consultation intelligent customer service system based on a large language model according to claim 1, characterized in that: The expression for checking whether the information is complete in the intelligent reply generation module is: , which means checking whether the generated response contains complete information; , indicating that if the information is incomplete , the user is prompted to provide additional information; , indicating that if the information is complete , then provide suggestions or other responses; The final reply expression is generated as: 。 9. The maintenance consultation intelligent customer service system based on a large language model according to claim 1, characterized in that: The optimization expression for the preliminary response in the response optimization and quality control module is: , : represents the initial response generated; : Indicates the final response after optimization and quality control.

Images