Customer service prompting system and method based on large language model

Abstract

The invention discloses a customer service prompting system and method based on a large language model, and relates to the field of customer service prompting technology research, and the system comprises a case collection module, a data cleaning module, a chat case processing module, a document segmentation processing module and a customer service prompting module. According to the system, cases are collected through a case collection module, irrelevant information in case data is cleaned through a data cleaning module, model training is carried out through a chat case processing module, semantic analysis and segmentation rules are formulated through a document segmentation processing module to carry out segmentation processing on a document, and the chat case processing module carries out chat case processing. Then, an application end in the customer service prompt module is used for obtaining problems of the user in real time, and suggestions for the problems of the customer are automatically generated through a large language model; according to the method, reply suggestions to customers are automatically generated for the questions in the aspect of psychological counseling proposed by the customers through the large language model, so that customer service personnel can select the reply suggestions, and the efficiency of customer service personnel when answering the customer questions and the personalization of question answering are improved.

Description

Technical Field

[0001] The present invention relates to the research field of customer service prompt technology, and in particular to a customer service prompt system and method based on a large language model. Background Art

[0002] With the continuous development of artificial intelligence technology, the large language model intelligent customer service system based on the knowledge base has become a popular choice for enterprises to provide customer service. Such systems can simulate human natural language communication capabilities through big data and deep learning technology, thereby providing customers with more efficient and personalized question-answering services. However, the current customer service prompt system based on artificial intelligence has the problem of untimely information updates, resulting in the inability to provide users with accurate and effective latest information when generating reply suggestions. In addition, in multiple rounds of conversations, it is easy to forget previous information, deviate from the topic, or repeat questions. In addition, since artificial intelligence cannot perceive the user's emotional state like human customer service, and the replies are relatively stiff and stereotyped, when faced with complex questions from psychological counseling users, it cannot provide effective guidance and follow-up questions, and the output reply suggestions are stereotyped and cannot soothe the user's emotions or solve problems. Summary of the invention

[0003] The purpose of the present invention is to provide a customer service prompt system and method based on a large language model to solve the problems raised in the prior art.

[0004] To achieve the above object, the present invention provides the following technical solutions: a customer service prompt system based on a large language model, the system comprising a case collection module, a data cleaning module, a chat case processing module, a document segmentation processing module and a customer service prompt module;

[0005] The case collection module collects and stores different types of cases by recording conversations between customer service and customers;

[0006] The data cleaning module cleans the collected case data by using a denoising method in natural language processing technology;

[0007] The chat case processing module formats all collected cases to form a data set, and inputs the data set into a large language model for model training;

[0008] The document segmentation processing module segments the document by formulating semantic analysis and segmentation rules, generates semantic segments and converts the semantic segments into high-dimensional vectors;

[0009] The customer service prompt module obtains user questions in real time through the application end and automatically generates suggestions for customer questions.

[0010] Furthermore, the data cleaning module includes an irrelevant information filtering unit, an error correction unit and a non-text information processing unit; the irrelevant information filtering unit is used to filter out meaningless conversation content in the collected cases to reduce data noise; the error correction unit uses a spelling checker tool to automatically retrieve spelling errors in the text; the non-text information processing unit is used to convert pictures and voice information into text information, and filter out emoticons in the text information.

[0011] Furthermore, the chat case processing module includes a chat case formatting unit and a model training unit; the chat case formatting unit formats all collected cases using a program; and the model training unit performs model training on the formatted data set through a large language model.

[0012] Furthermore, the document segmentation processing module uses natural language processing technology to locate the natural dividing points of the document and formulate semantic analysis and segmentation rules based on this, and segments the document according to the segmentation rules; after segmenting the document, independent semantic fragments are generated, and the segmented text fragments are converted into high-dimensional vectors using an embedding model. A high-dimensional vector is a numerical vector that is formed by converting a text composed of multiple sentences when text vectorization technology in natural language processing is used for text vectorization. The vectors of all text fragments are stored in a vector database, so that the system can quickly retrieve the required semantic content based on semantic similarity.

[0013] Furthermore, the customer service prompt module uses the application end to obtain the user's questions in real time, and uses the cloud-based large language model to analyze the user's questions and automatically generate reply suggestions.

[0014] A customer service prompt method based on a large language model comprises the following steps:

[0015] S1. Collect conversation cases between customer service and customers by recording historical conversations; after the customer grants permission, record historical conversations and collect conversation cases between customer service and customers;

[0016] S2. Clean the collected case data using the denoising method in natural language processing technology;

[0017] S3, formatting the cases and inputting the formatted cases into the large language model for model training;

[0018] S4, using natural language processing technology to segment and vectorize the documents in the dialogue library, and storing the vectors of the segmented text segments in a vector database;

[0019] S5. Use the application to obtain customer questions and automatically generate response suggestions through the cloud-based large language model.

[0020] Furthermore, in step S2: in the collected conversation data between customer service and customers, the original data often contains various noise information, which will affect the learning effect of the model. Therefore, it is necessary to ensure that the data is structured, clean, and easy for the large model to process text; the case data is cleaned by the denoising method in natural language processing technology, so as to provide pure and high-quality data for the large language model. The data cleaning content includes the following three aspects:

[0021] a. Filter out irrelevant information in the collected case conversations to reduce data noise and ensure that the model focuses on effective conversation content; meaningless text includes conversations without actual information, such as "haha", "oh", "hmm", etc.;

[0022] b. Use spell checking tools to automatically detect and correct spelling errors in text to ensure data consistency and standardization. After correcting spelling errors, the model can learn and understand the customer's intentions more accurately.

[0023] c. The original conversation data may contain non-text information such as pictures, voices, and emoticons. These contents cannot be directly input into the large model for training, so they need to be specially processed. The non-text content in the filtering and processing cases includes picture conversion, voice conversion, and emoticon filtering. Picture conversion uses OCR technology to extract text from pictures and insert it into the conversation text. Voice conversion uses voice recognition technology to transcribe the voice data input by the user into text and insert it into the conversation content. Emoticon filtering uses the symbol filtering program to filter out emoticons in the conversation to prevent these symbols from affecting text analysis.

[0024] Furthermore, in step S3: firstly, the cases are formatted, and then the formatted case data set is input into the large language model for model training. The model training steps are as follows:

[0025] S3-1. Load the formatted and structured JSON dialogue dataset into the training frameworks TensorFlow and PyTorch. The data is preprocessed to meet the input requirements of the model so that it can be effectively understood and processed by the model.

[0026] S3-2. Select a large language model architecture, such as GBT or BERT, and load the pre-trained basic model. Use the language understanding capabilities of the pre-trained model to fine-tune the model to the specific area of ​​customer service dialogue through transfer learning.

[0027] S3-3. Load the pre-trained weights during the model initialization phase to give the model general language understanding capabilities, so that the model can quickly acquire basic semantic understanding, and fine-tune the pre-trained model on the specific data of customer service conversations to gradually adapt to the semantic needs of this field; this process enables the model to transition from general understanding to specific understanding of customer service conversations;

[0028] S3-4. In the fine-tuning stage, the model is trained in a targeted manner using supervised learning methods. By repeatedly inputting these message pairs, the model gradually enhances its ability to understand the semantics of specific situations and generate accurate responses. Loss functions such as cross entropy are used to measure the gap between the model prediction and the true label, thereby guiding the optimization of model parameters.

[0029] S3-5. After fine-tuning the model, continuously optimize the training model through feedback; monitor the model performance in real time in actual conversations, add newly generated conversation data and customer service feedback to the training data set, and regularly retrain or fine-tune the model to enhance the model's adaptability and generation quality, so that it maintains efficient understanding and responsiveness in ever-changing conversation scenarios.

[0030] Furthermore, in step S4: the collected chat case documents are segmented and the segmented text segments are converted into high-dimensional vectors and stored in a vector database; the document segmentation and vectorization process includes the following steps:

[0031] S4-1. Use natural language processing technology to perform semantic analysis on documents to identify the theme and content of each paragraph, and formulate text segmentation rules according to logical segments and semantic coherence; for example, it can use sentence models or dependency syntactic analysis to find natural dividing points so that each segment has independent semantic unit characteristics;

[0032] S4-2. Divide the document into segments according to the set semantic logic. Each segment is independent and complete in meaning, so as to provide relevant responses quickly when customers ask questions. To optimize retrieval performance, each segment should contain core topic tags and keywords and maintain an appropriate length (usually 3-5 sentences) to avoid being too short or too long to affect the understanding effect, so as to ensure that the segment has the ability to express independently and be self-sufficient in information.

[0033] S4-3. Use the embedded model to convert the segmented text fragments into high-dimensional vectors, and store the vectors of all text fragments in the database; the embedding model captures the semantic features of the text and represents each fragment as a vector for semantic similarity calculation; after vectorization, the text is no longer a simple character or phrase, but is converted into a point in the vector space; in this space, the distance between vectors reflects the semantic similarity; the vectors of all text fragments are stored in the vector database, so that the system can quickly retrieve relevant content based on semantic similarity, thereby improving the retrieval accuracy and ensuring that the provided content has a higher degree of semantic match with the customer's question.

[0034] Further, in step S5: the user's question is obtained in real time through the application end, and the reply suggestion is automatically generated through the trained large language model. The customer service staff refers to the automatically generated reply suggestion to reply the user. For customers of psychological consultation, the customer service staff's purpose is to first understand the customer's general situation and the confusion encountered, and recommend a professional psychological consultant corresponding to the situation. The cloud-based large model is used to analyze the customer's problem and automatically generate a request reply for the customer's problem. The specific reply process includes the following steps:

[0035] S5-1, the application side obtains chat records in real time through the session archive interface, and stores and processes the chat records on its own server side;

[0036] S5-2: When a customer service representative or customer sends a message, the application captures the current conversation in real time, obtains all chat messages sent by the current customer and the current customer before, uses these contents as input information, and automatically requests the cloud-based large language model to analyze and process the chat messages;

[0037] S5-3. The big language model converts the customer's question into a vector, searches the vector database for vectorized segments that are similar to the customer's question and are ranked in the top five among all similar segments, and uses the Langchain method to pass these five vectorized texts to the big language model, which analyzes them and summarizes reasonable reply suggestions.

[0038] S5-4. The reply suggestions generated by the large language model are immediately fed back to the application end and displayed in the customer service staff's operation interface. After viewing the reply suggestions, the customer service staff decides whether to adopt the suggestions. If they choose to adopt them, the customer service staff can directly copy the suggestions and send them to the customer. If they do not choose to adopt them, the customer service staff can edit or write the reply suggestions themselves. Alternatively, the customer service staff can slightly modify the automatically generated reply suggestions and send a complete and accurate reply suggestion to the user.

[0039] S5-5. When the customer sends a new message, the application recaptures the new conversation and restarts the entire process, from conversation content analysis to generating the latest reply suggestions; this process continues until the entire conversation ends.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] In the system preparation stage, the present invention obtains a large number of chat case training models by collecting and processing massive chat cases, thereby solving the problems of timely updating and storage of system information. In addition, the application end captures and analyzes chat information in previous rounds of conversations in each round of conversation, thereby solving the logical coherence of the conversation with the customer automatically generated by the large language model. In addition, the present invention can guide and deeply explore user demands according to the questions input by the user. When consulting psychological counseling customers, the large language model can be used to analyze and generate reply suggestions to soothe the customer's emotions and address the customer's questions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a structural schematic diagram of a customer service prompt system based on a large language model of the present invention;

[0043] Figure 2 The figure is a flow chart of a customer service prompt method based on a large language model according to the present invention. DETAILED DESCRIPTION

[0044] 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.

[0045] like Figure 1-Figure 2 As shown, the present invention provides a technical solution, a customer service prompt system based on a large language model, the system includes a case collection module, a data cleaning module, a chat case processing module, a document segmentation processing module and a customer service prompt module;

[0046] The case collection module collects and stores different types of cases by recording conversations between customer service and customers;

[0047] The data cleaning module cleans the collected case data by using a denoising method in natural language processing technology;

[0048] The chat case processing module formats all collected cases to form a data set, and inputs the data set into a large language model for model training;

[0049] The document segmentation processing module segments the document by formulating semantic analysis and segmentation rules, generates semantic segments and converts the semantic segments into high-dimensional vectors;

[0050] The customer service prompt module obtains user questions in real time through the application end and automatically generates suggestions for customer questions.

[0051] The data cleaning module includes an irrelevant information filtering unit, an error correction unit and a non-text information processing unit; the irrelevant information filtering unit is used to filter out meaningless conversation content in the collected cases to reduce data noise; the error correction unit uses a spelling checker tool to automatically retrieve spelling errors in the text; the non-text information processing unit is used to convert pictures and voice information into text information, and filter out emoticons in the text information.

[0052] The chat case processing module includes a chat case formatting unit and a model training unit; the chat case formatting unit formats all collected cases using a program; and the model training unit performs model training on the formatted data set through a large language model.

[0053] The document segmentation processing module uses natural language processing technology to locate the natural dividing points of the document and formulate semantic analysis and segmentation rules based on this, and segments the document according to the segmentation rules; after segmenting the document, independent semantic segments are generated, and the segmented text segments are vectorized using an embedding model, and the vectors of all text segments are stored in a vector database, so that the system can quickly retrieve the required semantic content based on semantic similarity.

[0054] The customer service prompt module uses the application end to obtain the user's questions in real time, and uses the cloud-based large language model to analyze the user's questions and automatically generate reply suggestions.

[0055] A customer service prompt method based on a large language model comprises the following steps:

[0056] S1. Collect conversation cases between customer service and customers by recording historical conversations; after the customer grants permission, record historical conversations and collect conversation cases between customer service and customers;

[0057] S2. Clean the collected case data using the denoising method in natural language processing technology;

[0058] S3, formatting the cases and inputting the formatted cases into the large language model for model training;

[0059] S4, using natural language processing technology to segment and vectorize the documents in the dialogue library, and storing the vectors of the segmented text segments in a vector database;

[0060] S5. Use the application to obtain customer questions and automatically generate response suggestions through the cloud-based large language model.

[0061] In step S2: In the collected conversation data between customer service and customers, the original data often contains various noise information, which will affect the learning effect of the model. Therefore, it is necessary to ensure that the data is structured, clean, and easy for the large model to process text; the case data is cleaned by the denoising method in natural language processing technology to provide pure and high-quality data for the large language model. The data cleaning content includes the following three aspects:

[0062] a. Filter out irrelevant information in the collected case conversations to reduce data noise and ensure that the model focuses on effective conversation content; meaningless text includes conversations without actual information, such as "haha", "oh", "hmm", etc.;

[0063] b. Use spell checking tools to automatically detect and correct spelling errors in text to ensure data consistency and standardization. After correcting spelling errors, the model can learn and understand the customer's intentions more accurately.

[0064] c. The original conversation data may contain non-text information such as pictures, voices, and emoticons. These contents cannot be directly input into the large model for training, so they need to be specially processed. The non-text content in the filtering and processing cases includes picture conversion, voice conversion, and emoticon filtering. Picture conversion uses OCR technology to extract text from pictures and insert it into the conversation text. Voice conversion uses voice recognition technology to transcribe the voice data input by the user into text and insert it into the conversation content. Emoticon filtering uses the symbol filtering program to filter out emoticons in the conversation to prevent these symbols from affecting text analysis.

[0065] In step S3: first, the case is formatted, and then the formatted case data set is input into the large language model for model training. The model training steps are as follows:

[0066] S3-1. Load the formatted and structured JSON dialogue dataset into the training frameworks TensorFlow and PyTorch. The data is preprocessed to meet the input requirements of the model so that it can be effectively understood and processed by the model.

[0067] S3-2. Select a large language model architecture, such as GBT or BERT, and load the pre-trained basic model. Use the language understanding capabilities of the pre-trained model to fine-tune the model to the specific area of ​​customer service dialogue through transfer learning.

[0068] S3-3. Load the pre-trained weights during the model initialization phase to give the model general language understanding capabilities, so that the model can quickly acquire basic semantic understanding, and fine-tune the pre-trained model on the specific data of customer service conversations to gradually adapt to the semantic needs of this field; this process enables the model to transition from general understanding to specific understanding of customer service conversations;

[0069] S3-4. In the fine-tuning stage, the model is trained in a targeted manner using supervised learning methods. By repeatedly inputting these message pairs, the model gradually enhances its ability to understand the semantics of specific situations and generate accurate responses. Loss functions such as cross entropy are used to measure the gap between the model prediction and the true label, thereby guiding the optimization of model parameters.

[0070] S3-5. After fine-tuning the model, continuously optimize the training model through feedback; monitor the model performance in real time in actual conversations, add newly generated conversation data and customer service feedback to the training data set, and regularly retrain or fine-tune the model to enhance the model's adaptability and generation quality, so that it maintains efficient understanding and responsiveness in ever-changing conversation scenarios.

[0071] In step S4: the collected chat case documents are segmented and the segmented text segments are converted into high-dimensional vectors and stored in a vector database; the document segmentation and vectorization process includes the following steps:

[0072] S4-1. Use natural language processing technology to perform semantic analysis on documents to identify the theme and content of each paragraph, and formulate text segmentation rules according to logical segments and semantic coherence; for example, it can use sentence models or dependency syntactic analysis to find natural dividing points so that each segment has independent semantic unit characteristics;

[0073] S4-2. Divide the document into segments according to the set semantic logic. Each segment is independent and complete in meaning, so as to provide relevant responses quickly when customers ask questions. To optimize retrieval performance, each segment should contain core topic tags and keywords and maintain an appropriate length (usually 3-5 sentences) to avoid being too short or too long to affect the understanding effect, so as to ensure that the segment has the ability to express independently and be self-sufficient in information.

[0074] S4-3. Use the embedded model to convert the segmented text fragments into high-dimensional vectors, and store the vectors of all text fragments in the database; the embedding model captures the semantic features of the text and represents each fragment as a vector for semantic similarity calculation; after vectorization, the text is no longer a simple character or phrase, but is converted into a point in the vector space; in this space, the distance between vectors reflects the semantic similarity; the vectors of all text fragments are stored in the vector database, so that the system can quickly retrieve relevant content based on semantic similarity, thereby improving the retrieval accuracy and ensuring that the provided content has a higher degree of semantic match with the customer's question.

[0075] In step S5: the user's question is obtained in real time through the application end, and the reply suggestion is automatically generated through the trained large language model. The customer service staff refers to the automatically generated reply suggestion to reply the user. For customers of psychological consultation, the customer service staff's purpose is to first understand the customer's general situation and the confusion encountered, and recommend a professional psychological consultant corresponding to the situation. The cloud-based large model is used to analyze the customer's problem and automatically generate a request reply for the customer's problem. The specific reply process includes the following steps:

[0076] S5-1, the application side obtains chat records in real time through the session archive interface, and stores and processes the chat records on its own server side;

[0077] S5-2: When a customer service representative or customer sends a message, the application captures the current conversation in real time, obtains all chat messages sent by the current customer and the current customer before, uses these contents as input information, and automatically requests the cloud-based large language model to analyze and process the chat messages;

[0078] S5-3. The big language model converts the customer's question into a vector, searches the vector database for vectorized segments that are similar to the customer's question and are ranked in the top five among all similar segments, and uses the Langchain method to pass these five vectorized texts to the big language model, which analyzes them and summarizes reasonable reply suggestions.

[0079] S5-4. The reply suggestions generated by the large language model are immediately fed back to the application end and displayed in the customer service staff's operation interface. After viewing the reply suggestions, the customer service staff decides whether to adopt the suggestions. If they choose to adopt them, the customer service staff can directly copy the suggestions and send them to the customer. If they do not choose to adopt them, the customer service staff can edit or write the reply suggestions themselves. Alternatively, the customer service staff can slightly modify the automatically generated reply suggestions and send a complete and accurate reply suggestion to the user.

[0080] S5-5. When the customer sends a new message, the application recaptures the new conversation and restarts the entire process, from conversation content analysis to generating the latest reply suggestions; this process continues until the entire conversation ends.

[0081] Embodiment 1:

[0082] When encountering consultation information sent by customers through the enterprise WeChat application, the system immediately starts to collect and analyze the customer's questions through the application, and automatically generates reply suggestions for customer service through the cloud-based large language model for reference; and for psychological consultation customers, the customer service staff's goal is to first understand the customer's general situation and the confusion they encounter, and then recommend a professional psychological consultant corresponding to their situation; the following is an example of an actual usage scenario:

[0083] Customer: "Hello! My relationship with my husband is not good and we often quarrel."

[0084] The application automatically sends the customer's sentence to the cloud model to obtain request response suggestions.

[0085] The cloud-based big language model first analyzes the meaning of this sentence and concludes that the client is confused about her marriage and emotions, but this is not enough to further determine the specific type of marriage and emotional problem.

[0086] At this time, the cloud returns a reply suggestion and displays it on the application interface: "I am very concerned about your current situation. Now you can slowly tell me some details about your relationship as a couple and what you often quarrel about. We are not in a hurry. There is plenty of time. You can talk slowly. I am willing to listen to you."

[0087] After seeing the reply suggestion, the customer service staff decided to adopt it and then sent the entire reply to the customer.

[0088] Customer service: "I am very concerned about your current situation. Can you tell me some details about your relationship and why you often quarrel? We are not in a hurry. We have plenty of time. You can talk slowly. I am willing to listen to you."

[0089] Client: "My husband doesn't care about the family. He is busy with work and comes home very late. He doesn't care about anything at home. I take care of the children at home. I don't mind him doing housework, which is fine with me, but he doesn't care at all. He even forgot the child's birthday a few days ago. The child and I waited at home for him to eat cake, but he didn't come home until the child fell asleep. He came back at 11 o'clock at night, and I complained to him a few words, and then we had a quarrel."

[0090] The application sends all the previous conversations between the customer and the customer service to the cloud, and again requests a reply suggestion, and gets: "Do you often encounter such a scenario? Have you sat down with him to have a good chat about this situation?"

[0091] Customer service: "Do you often encounter such situations? Have you sat down with him to have a good chat about this situation?"

[0092] Customer: "I want to talk too, but every time I try to communicate, we end up arguing after just a few words. It always ends in a bad mood."

[0093] The application again requests a reply suggestion from the cloud. At this point, the large model (or small model) has determined that the client is a type with communication barriers between husband and wife, and provides a reply suggestion that leads to the result: "I really feel sorry for you! I think you must be very aggrieved at this time? You want to live a good life, but you are powerless, right? In fact, the main problem between you is communication. There are always problems in the communication between husband and wife, which makes everything else unable to proceed normally. How about this, I have a psychological consultant who is very good at solving communication problems. She has solved hundreds of similar situations. I will send you her WeChat. You add her and she will give you a detailed analysis and some suggestions. I will send her all the situations we talked about above, so you don’t need to repeat them. Is this okay with you?"

[0094] The customer service staff will forward the above reply suggestions to the customer and assist the customer in connecting with the psychological counselor to ensure smooth subsequent communication.

[0095] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A customer service prompt system based on a large language model, characterized in that: The system includes case collection module, data cleaning module, chat case processing module, document segmentation processing module and customer service prompt module; The case collection module collects and stores different types of cases by recording conversations between customer service and customers; The data cleaning module cleans the collected case data by using a denoising method in natural language processing technology; The chat case processing module formats all collected cases to form a data set, and inputs the data set into a large language model for model training; The document segmentation processing module segments the document by formulating semantic analysis and segmentation rules, generates semantic segments and converts the semantic segments into high-dimensional vectors; The customer service prompt module obtains user questions in real time through the application end and automatically generates suggestions for customer questions.

2. A customer service prompt system based on a large language model according to claim 1, characterized in that: The data cleaning module includes an irrelevant information filtering unit, an error correction unit and a non-text information processing unit; the irrelevant information filtering unit is used to filter out the conversation contents that have no practical significance in the collected cases; The error correction unit uses a spelling checker to automatically retrieve spelling errors in the text; the non-text information processing unit is used to convert image and voice information into text information and filter out emoticons in the text information.

3. A customer service prompt system based on a large language model according to claim 1, characterized in that: The chat case processing module includes a chat case formatting unit and a model training unit; the chat case formatting unit formats all collected cases; and the model training unit performs model training on the formatted data set through a large language model.

4. The customer service prompt system based on a large language model according to claim 1, characterized in that: The document segmentation processing module uses natural language processing technology to locate the natural dividing points of the document and formulate semantic analysis and segmentation rules based on them, and segments the document according to the segmentation rules; after segmenting the document, independent semantic segments are generated, and the segmented text segments are vectorized using an embedding model, and the vectors of all text segments are stored in a vector database.

5. The customer service prompt system based on a large language model according to claim 1, characterized in that: The customer service prompt module uses the application end to obtain the user's questions in real time, and uses the cloud-based large language model to analyze the user's questions and automatically generate reply suggestions.

6. A customer service prompt method based on a large language model, characterized in that: The method comprises the following steps: S1. Collect conversation cases between customer service and customers by recording historical conversations; S2. Clean the collected case data using the denoising method in natural language processing technology; S3, formatting the cases and inputting the formatted cases into the large language model for model training; S4, using natural language processing technology to segment and vectorize the documents in the dialogue library, and storing the vectors of the segmented text segments in a vector database; S5. Use the application to obtain customer questions and automatically generate response suggestions through the cloud-based large language model.

7. The customer service prompt method based on a large language model according to claim 6, characterized in that: In step S2: the case data is cleaned by using the denoising method in natural language processing technology. The data cleaning content includes the following three aspects: a. Filter out meaningless conversations from the collected case conversations; b. Use spell-checking tools to automatically detect and correct spelling errors in text; c. Filtering and processing non-text content in cases includes image conversion, voice conversion, and emoticon filtering; image conversion uses OCR technology to extract text from images and insert it into the conversation text; Voice conversion uses speech recognition technology to transcribe the user's input voice data into text and insert it into the conversation content; Emoji Filter Filter out emojis from your conversations with the Emoji Filter.

8. The customer service prompt method based on a large language model according to claim 6, characterized in that: In step S3: first, the case is formatted, and then the formatted case data set is input into the large language model for model training. The model training includes the following steps: S3-1. Load the dialogue dataset formatted and structured in JSON format into the training framework; S3-2. Load the pre-trained basic model through the large language model architecture, and use the language understanding ability of the pre-trained model to fine-tune the model to the specific field of customer service dialogue through transfer learning; S3-3. Load the pre-trained weights during the model initialization phase to give the model general language understanding capabilities, and fine-tune the pre-trained model on specific data from customer service conversations; S3-4. In the fine-tuning stage, supervised learning methods are used to train the model in a targeted manner; S3-5. Monitor the model performance in real time after fine-tuning the model, add the newly generated conversation data and customer service feedback to the training data set, and retrain the model regularly based on the feedback.

9. The customer service prompt method based on a large language model according to claim 6, characterized in that: In step S4: segmenting the collected chat case documents and converting the segmented text segments into high-dimensional vectors and storing them in a vector database; The process of document segmentation and vectorization includes the following steps: S4-1. Use natural language processing technology to perform semantic analysis on documents to identify the theme and content of each paragraph, and formulate text segmentation rules according to logical segments and semantic coherence; S4-2, divide the document into segments according to the set semantic logic, and each segment exists independently and has complete meaning; S4-3. Use the embedded model to convert the segmented text segments into high-dimensional vectors, and store the vectors of all text segments in the database.

10. The customer service prompt method based on a large language model according to claim 6, characterized in that: In step S5: the user's question is obtained in real time through the application end, and reply suggestions are automatically generated through the trained large language model. The customer service staff responds to the user with reference to the automatically generated reply suggestions. The specific reply process includes the following steps: S5-1, the application side obtains chat records in real time through the session archive interface, and stores and processes the chat records on its own server side; S5-2: When a customer service representative or customer sends a message, the application captures the current conversation in real time, obtains all chat messages sent by the current customer and the current customer before, uses these contents as input information, and automatically requests the cloud-based large language model to analyze and process the chat messages; S5-3. The big language model converts the customer's question into a vector, searches the vector database for vectorized segments that are similar to the customer's question and are ranked in the top five among all similar segments, and uses the Langchain method to pass these five vectorized texts to the big language model, which analyzes them and summarizes the reply suggestions. S5-4, the large language model generates a reply suggestion and then feeds it back to the application end, and displays it in the customer service staff's operation interface; the customer service staff decides whether to adopt the reply suggestion after viewing it; S5-5. When the customer sends a new message, the application recaptures the new conversation and restarts the entire process, from conversation content analysis to generating the latest reply suggestions; this process continues until the entire conversation ends.

Images