Bank intelligent customer service question answering method and system based on large model

Abstract

The invention discloses a bank intelligent customer service question answering method and system based on a large model, relates to the technical field of intelligent question answering, and solves the technical problem that although the advantages of the large model are utilized in the prior art, the large model needs to answer according to comprehensive information, and the use scene is limited. The method comprises the steps that a consultation request sent by a user through a client is received, a standard answer is returned according to cosine similarity, if no standard answer exists, a question is routed according to an execution result of a group script, index processing is carried out on data sent by an initial upstream system through the group script of index processing, and the index processing is carried out on the data sent by the upstream system. A cue word template is determined and sent to the large model, if a result returned by the large model has a specific symbol, corresponding information needs to be inquired and sent to the large model, the large model thinking and replying again and returns the result to the client, and the operation is repeated until the user actively ends inquiry; according to the invention, intelligent response support for questions of clients is realized, accurate and compliant reply is ensured, the system logic is flexible and controllable, and the reply personification degree is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent question answering, and in particular to a bank intelligent customer service question answering method and system based on a large model. BACKGROUND

[0002] The existing technical solutions generally have two kinds of question and answer pair matching solutions based on FAQ and prompt word routing solutions based on large models.

[0003] (1) The question and answer pair matching solution based on FAQ requires that a large number of customer questions and corresponding answers to the questions be entered in the form of QA pairs in advance and maintained in the system. When a user raises a question, the closest question in the historical QA pairs is found through certain rules, and the answer to the user's question is returned. The defect is that it relies on system maintenance personnel to collect a large number of standard questions for import. The manual cost of maintaining the question library is high, and this type of solution can only have a single round of dialogue. To implement multi-round dialogue, a more complex neural network model needs to be introduced. Moreover, since the system's replies are all from standard answers, they are very rigid and have low anthropomorphism.

[0004] For example, Chinese patent "Intelligent customer service implementation method and system based on deep semantic learning" (Patent application number: CN202310838313.3, publication number: CN116561287B). The patent mentions that a deep learning model can be used to classify user questions by text, and see which category of existing question library the user question can be classified into. If it is classified into an existing question, the standard answer to the question is returned. However, this solution needs to regularly train new deep neural networks to cover newly entered questions, which has poor flexibility.

[0005] For example, Chinese patent "Response method and device of robot, electronic equipment and storage medium" (Patent application number: CN202411602860.2, publication number: CN119691105A). The patent introduces a vector database, and all collected standard questions are converted into multi-dimensional vectors and stored in the vector database in advance. After the user raises a question, the user's question is also converted into a vector and the high-performance vector query algorithm (generally hierarchical small-world navigation algorithm HNSW) of the vector database is used to quickly query the standard answer to the user's question. This solution can add new standard questions to the system at any time, and has improved flexibility, but still relies on system maintenance personnel to collect a large number of standard questions for import.

[0006] (2) Based on the large model prompt word routing scheme, after the large language model (LLM) technology matures, this technology is also applied to customer service systems. The large language model and the traditional FAQ question and answer pair are different. The large language model itself has a large amount of knowledge that can answer various questions. In order to improve the accuracy of the large model answer, a routing method is designed. However, the large model-based scheme has obvious defects, including: 1. Due to the randomness and uncontrollability of the large model answer, especially the large model that has been fine-tuned, it is likely to cause catastrophic forgetting. For some red line issues of financial regulation, if the model gives a wrong answer, it will cause extremely serious consequences. 2. The previous scheme needs to query all kinds of information of the user as much as possible at the beginning of the dialogue, and assemble the speech into a speech template through the corresponding prompt word, and send it to the large model. The large model answers according to the comprehensive information. This greatly limits the use scene. In the past, when relying entirely on human customer service, the data collected was incomplete, and many times it relied on the active inquiry of human customer service and queried external interfaces to supplement the information. The existing method cannot cover the scene that needs the large model to actively inquire and collect information. (For example, when the user asks about my available limit, if we do not have the user's mobile phone number or ID number, we cannot query it. In the era of human customer service, human customer service will actively ask the user's mobile phone number or ID number, and log in to the corresponding system to query the available limit of the customer)

[0007] For example, Chinese patent "Question information processing method, device, equipment, storage medium and product" (Patent application number: CN202311470635.3, publication number: CN117453885A). The patent proposes a question information routing scheme, which distributes the user's question to the corresponding large language model by judging whether it belongs to a professional field or a general field, in order to improve the accuracy of the answer, but this patent needs to invest a lot of hardware resources to train multiple large language models in professional fields, which is very costly.

[0008] For example, Chinese patent "Large model prompt word intelligent routing method, device, equipment and storage medium" (Patent application number: CN202510206990.2, publication number: CN119691137A). The patent discloses a large model prompt word intelligent routing method, which generates a decision tree based on a code file and a preset database to select different knowledge models to process user prompt words, but introduces a relatively cumbersome decision engine system. SUMMARY

[0009] In order to solve the problems existing in the prior art, the application provides a bank intelligent customer service question and answer method and system based on a large model, which solves the technical problems that although the prior art utilizes the advantages of a large model, the large model needs to answer according to comprehensive information, the use scene is limited, and the catastrophic forgetting of the large model may touch the hidden danger of the red line of financial supervision.

[0010] A bank intelligent customer service question and answer method based on a large model, comprising:

[0011] Step 1: judging whether the consultation question of the customer is a regulatory red line question, if yes, directly returning a standard answer, if not, entering step 2;

[0012] Step 2: routing the question to a specified prompt word template according to the key words in the question;

[0013] Step 3: performing index processing on the received data to obtain prompt word parameters required by the prompt word template;

[0014] Step 4: assembling the prompt word parameters to the specified prompt word template to obtain a complete speech to be sent to the large model, and obtaining the result returned by the large model;

[0015] Step 5: if the result returned by the large model exists a specific symbol, querying corresponding information from a user database according to the following content of the specific symbol and returning the information to the large model, and the large model again thinks and answers, if the specific symbol does not exist, directly returning the result returned by the large model to the customer;

[0016] Step 6: repeating steps 1-5 until the user actively ends the inquiry.

[0017] Further, the step 1 comprises

[0018] Step 1.1: receiving a consultation request sent by a user through a client, containing a question string sent by the user and user basic information;

[0019] Step 1.2: taking out the question string of the user, and converting the question string into a high-dimensional vector through a text-to-vector model;

[0020] Step 1.3: searching the high-dimensional vector obtained in step 1.2 in a vector database, selecting the closest question, and judging whether the cosine similarity between the two is higher than a threshold value, if yes, directly returning a standard answer, if not, entering step 2, the vector database saving regulatory red line questions;

[0021] Further, the step 1.2 comprises:

[0022] Step 1.2.1: obtaining the question string of the user;

[0023] Step 1.2.2: Remove redundant spaces, special characters that have no effect on semantics;

[0024] Step 1.2.3: The processed string is sent to the text-to-vector model to obtain a high-dimensional vector.

[0025] Further, the prompt word template further includes large model parameters, including large model type, large model temperature, system, and maximum token.

[0026] Further, the step 3 includes: performing index processing on the received data to obtain the prompt word parameters required by the prompt word template; performing index processing on the original parameters sent by the initial upstream system to obtain processed parameters, the index processing including structured processing and numerical calculation of data, and determining the prompt word parameters from the processed parameters according to the requirements of the prompt word template.

[0027] Further, steps 2 and 3 are implemented by groovy scripts, and when the groovy scripts are changed during execution, the groovy scripts are compiled first and then loaded into memory.

[0028] Further, the step 4 includes:

[0029] Step 4.1: The first dialogue generates an opening speech according to the prompt word template for the large model;

[0030] Step 4.2: The subsequent dialogue directly sends the user's question to the large model.

[0031] Further, if the customer is in the visitor mode, the prompt word template received by the large model is missing key parameters including mobile phone number and name, the large model will inquire the key information one by one, and call external systems to obtain information to verify whether the customer's reply is corresponding, if corresponding, the large model will answer the question according to the customer's demand, if not corresponding, the large model will ask the customer whether the information is wrong, if the customer's reply is not corresponding for multiple times, the large model will judge that the customer has no query right, and stop the dialogue after issuing a closing speech.

[0032] A bank intelligent customer service question and answer system based on a large model, comprising: a client, a system, a vector database, and a large model port, the client is used for customer interaction with the system, the system is used for executing a bank intelligent customer service question and answer method based on a large model, the vector database is used for saving high-dimensional vectors of existing questions and standard answers, and the large model port is used for receiving system instructions and calling corresponding large models to reply to questions and return results.

[0033] The beneficial effects of the present application include:

[0034] 1. On the basis of a large model, vector semantic retrieval is introduced to intercept the red line problem of financial supervision. Compared with the past FAQ-based scheme, the number of QAs in this proposal is greatly reduced, and only a few sensitive problems need to be concerned. But it can well solve the uncontrollability after the catastrophic forgetting in the past large model-based scheme, effectively reducing the customer complaint rate.

[0035] 2. Introduce groovy code to dynamically process prompt word routing. Groovy code can be compiled and loaded at runtime, and has good flexibility as the decision engine, but groovy is more lightweight and does not need to introduce large systems. At the same time, compared with the scheme using large model routing, the resource consumption of groovy is negligible.

[0036] 3. Large model active inquiry mechanism, if the customer lacks necessary information, the large model can actively initiate inquiry, and extract and complete user information according to user answers, or inform the system of the invention according to the extracted information to trigger additional queries, greatly improving the personification, and covering more comprehensive scenarios, which can cope with various customer service, marketing, electric stimulation and other scenarios. In the case of missing information, inaccurate customer information and other situations, it can be handled, compared with the past scheme that can only be transferred to manual, the use effect is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A large model-based bank intelligent customer service question and answer method process related to an embodiment of the present application Figure 1 .

[0038] Figure 2 A large model-based bank intelligent customer service question and answer method process related to an embodiment of the present application Figure 2 .

[0039] Figure 3 A large model-based bank intelligent customer service question and answer system block diagram related to an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Therefore, the detailed description of the embodiments of the present application provided in the following drawings is not intended to limit the scope of the claimed present application, but only to represent selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] Embodiment 1

[0042] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification. Figures 1-3 The application will be described in greater detail with reference to specific embodiments.

[0043] A bank intelligent customer service question and answer method based on a large model, as shown in Figure 1 comprises the following steps:

[0044] Step 1: Determine whether the customer's consultation question is a regulatory red line question. If yes, return the standard answer directly. If not, go to step 2.

[0045] Step 2: Route the question to the specified prompt word template according to the keywords in the question.

[0046] Step 3: Process the received data to obtain the prompt word input required by the prompt word template.

[0047] Step 4: Assemble the prompt word input into the specified prompt word template to obtain a complete utterance for the large model, and obtain the result returned by the large model.

[0048] Step 5: If the result returned by the large model contains a specific symbol, query the corresponding information from the user database according to the following content of the specific symbol and return it to the large model. The large model will think again and reply. If there is no specific symbol, the result returned by the large model will be directly returned to the customer.

[0049] Step 6: Repeat steps 1-5 until the user initiates the end of the inquiry.

[0050] In another embodiment, a bank intelligent customer service question and answer method based on a large model, as shown in Figure 2 comprises the following steps:

[0051] Step 1: Receive the consultation request sent by the user through the client, including the question string sent by the user, user basic information, etc.

[0052] Example 1.1: The received request message is {"question":"How to query my available loan amount?", "phone":null, "channel":"WeChat", "name":"Zhang San"}, from which it can be determined that the user may be a visitor at this time and does not know his mobile phone number, ID number information.

[0053] Example 1.2: The received request message is {"question":"I am a student, can I apply for a credit loan?", "phone":null, "channel":"WeChat", "name":"Li Si"}, similarly, the mobile phone number and ID number information are unknown.

[0054] Step 2: Take out the user's question string and convert it into a high-dimensional vector through a text-to-vector model;

[0055] Step 2.1: Obtain the user's question string;

[0056] Step 2.2: Remove unnecessary spaces and special characters that have no effect on semantics;

[0057] Step 2.3: The processed string is sent to the text-to-vector model to obtain a high-dimensional vector.

[0058] Example 2.1: The sentence "How can I check my available loan amount?" may be converted into a high-dimensional vector consisting of pure doubles, such as [0.16, 0.31, 0.4812...].

[0059] Example 2.2: The sentence "Can I get a credit loan as a student?" may be converted into a high-dimensional vector consisting of pure doubles, such as [0.7, 0.64, 0.12...].

[0060] Step 3: Retrieve the high-dimensional vector obtained in Step 2 in the vector database, select the closest question, and determine if the cosine similarity is higher than the threshold. If it is, return the standard answer directly. If it is not, go to Step 4. The problems saved in the vector database are regulatory red line problems;

[0061] Example 3: Suppose the database retains the pre-set regulatory red line problem "Can a student get a credit loan?" and the corresponding vector [0.71, 0.63, 0.11...] is retained in the vector database. After similarity calculation, it is found that the vector similarity in example 2.1 is 0.1, which does not trigger the pre-set standard answer. The vector similarity in example 2.2 is 0.9, which is higher than the threshold, and the standard answer to this question is returned directly: "Hello, according to financial regulatory laws and regulations, it is not allowed to issue credit loans to students."

[0062] Step 4: Route the question to the specified prompt template according to the keywords in the question. After intercepting the regulatory red line problem in Step 3, routing will be performed according to the processing logic. Most routing is only performed at the beginning of this round of dialogue, and subsequent user questions are directly submitted to the large model.

[0063] In addition, the settings of the large model are also determined when routing the prompt template. The prompt template includes the selected model type, model temperature, system, maximum token, etc. These parameters will affect the flexibility and conciseness of the model's answers to questions, etc.

[0064] Example 4: If the user question contains the keyword "loan", route to the prompt word template: "The user is consulting a question through {channel_zh}: {question}, please play the role of a loan consulting manager of XX Bank, and briefly answer the user's question"; if the user question contains the keyword "deposit", route to the prompt word template "The user is consulting a deposit-related question through {channel_zh}: {question}, please answer the user's question briefly in combination with the actual business of XX Bank and the interest rate situation". Then example 3 will be routed to the first prompt word template.

[0065] Step 5: Index processing is performed on the received data to obtain the prompt word parameters required by the prompt word template; through the groovy script of index processing, the original parameters sent by the initial upstream system are processed to obtain the processed parameters, which includes structured processing and calculation of data, conversion of numbers to Chinese, and obtaining of the prompt word template in step 4, and determining the prompt word parameters from the processed parameters according to the requirements of the prompt word template. Similarly, this step is also only performed at the beginning of the current round of dialogue in most scenarios.

[0066] At present, the scale of large models ranges from tens of billions to hundreds of billions, and the smaller ones have 70-320 billion parameter specifications. Deepseek, which is good at mathematics and programming on the Internet, has more than 640 billion parameters. Different model types have different processing capabilities and response times. According to the scenario, if it is used for voice telephone, it tends to use small models below 32B, because the language telephone delay has a great impact on user experience, and the user needs to respond within a few hundred milliseconds after speaking. If it is a text-based customer service, the user's tolerance for delay is significantly improved, and a response within tens of seconds is acceptable. When there is sufficient computing power, even the deepseek complete version can be used.

[0067] If a large model with a smaller scale is used, the defect of such a large model is that the numerical processing and recognition ability is weak, and it is not suitable for too complex processing logic, and even when recognizing pure numbers, there will be some errors. At the beginning of the dialogue, the upstream system will send the customer's basic information together. Because the parameter of the small model is not good at mathematical calculation, in the field of banks, the parameters sent may include the customer's registration time or the last repayment time, etc. Therefore, this embodiment adopts calculation processing before sending the parameters into the large model, and the calculation results include how many days the last repayment is away from now and how many days the registration is, and these information will be sent to the large model through the prompt word template, to enhance the understanding of the large model for this customer, and the answer quality will be higher. By processing before the large model, the efficiency of question and answer is improved while ensuring the quality of question and answer when using a large model with a smaller parameter.

[0068] If a large-scale large model is adopted, although its processing ability is somewhat improved compared with the previously mentioned smaller-scale large model, which is manifested in longer conversation turns and support for more complex prompt words, there are still problems with unreliable calculation results. Facing complex date calculations, such as considering leap years, or calculating next Tuesday based on the current date, or repaying a loan after National Day, the reaction effect may not be ideal and is prone to errors. Therefore, in order to ensure the processing effect, it is still necessary to uniformly process the received data by processing indicators, process the information well, and then hand the accurate information to the large model.

[0069] Example 5: According to the description in Example 4, the prompt word template requires two input parameters: channel_zh and question, where question is directly provided by the upstream, but channel_zh needs to be processed. Suppose the pseudo-code logic for processing is as follows:

[0070] if channel == "WeChat" then channel_zh = "微信"

[0071] else channel_zh = "未知渠道"

[0072] Through this code, the input parameter channel_zh = "微信" can be obtained.

[0073] The parameters of different templates are not necessarily the same. For example, the prompt word template in the loan field may require parameters such as the customer's interest rate and amount. If it is the deposit field, using information such as the customer's current balance and fixed-term interest rate, after routing the customer to different templates, the prompt word template will not give all the processed parameters of the customer to the large model, but will select or process the required indicator parameters according to what the customer specifically wants to do, and the attention of the large model will be focused on the core goal.

[0074] Step 6: Assemble the prompt word input parameters into the specified prompt word template, obtain the complete utterance and send it to the large model, and obtain the result returned by the large model;

[0075] In another embodiment, if the conversation is initiated by the bank, the conversation is routed to the specified prompt word template according to the demand keyword and the customer's basic information. For example, when the question-and-answer system initiates a collection conversation, different-toned prompt word templates are selected according to the customer's debt duration.

[0076] In another embodiment, Steps 4 and 5 are implemented using groovy scripts. Therefore, when executing, if the groovy script changes, first compile the groovy script and then load it into memory; call the groovy script to route the prompt words.

[0077] The groovy script has the advantages of flexibility and modification at any time. Business personnel can adjust the routing strategy as needed without recompiling and packaging like java code, which is more suitable for different personnel to use.

[0078] In another embodiment, the step 6 comprises:

[0079] Step 6.1: The first dialogue generates an opening speech according to the prompt word template for the large model.

[0080] Step 6.2: The subsequent dialogue directly sends the user's question to the large model.

[0081] Example 6: The complete prompt word template is obtained and sent to the large model: "The user is consulting the question through WeChat: how to query my available loan amount? Please act as a loan consulting manager of XX bank, and briefly answer the user's question"

[0082] The next possible reply from the large model is: "You can query on the APP of my bank, the specific operation is XXXX".

[0083] Step 7: If the result returned by the large model contains a specific symbol, query the corresponding information from the user database according to the follow-up content of the specific symbol and return it to the large model, and the large model will answer again. If there is no specific symbol, the result returned by the large model will be directly returned to the client.

[0084] Example 7: For example, the large model has completed the following dialogue through the previous example 6:

[0085] To the large model (initial): "The user is consulting the question through WeChat: how to query my available loan amount? Please act as a loan consulting manager of XX bank, and briefly answer the user's question"

[0086] The reply of the large model: "You can query on the APP of my bank, the specific operation is XXXX"

[0087] The user continues to ask: "I am not convenient to operate now, can you help me query?"

[0088] The reply of the large model: "Yes, but I haven't received your mobile phone number, can you provide it?"

[0089] The user answers: "Yes, my mobile phone number is: 135xxxxx"

[0090] The large model replies to the special symbol for our system: "<query_credit>, arg: [phone = 135xxxx]"

[0091] Note that at this time the system will detect this special symbol <query_credit> and understand that this message is not to be given to the user, but the large model has delegated us to query, at this time the system will query other systems in the line according to the parameter phone=135xxxx to obtain the user's credit limit, assuming that the credit limit is credit=150000, at this time the system will tell the large model "the user's credit limit is 150000".

[0092] After the large model synthesizes the information given by our system, it replies: "Hello, we have found that your credit limit is 150000 yuan"

[0093] The user replies: "OK, thank you"

[0094] The conversation ends.

[0095] Step 8: Repeat steps 2-7 until the user voluntarily ends the inquiry, which includes two cases: the first is that the customer does not reply within a set time, and the second is that the customer sends information including keywords such as goodbye to indicate the end.

[0096] The other systems in the line refer to external systems that the large model system can access through interfaces. For example, a customer who is a tourist mode login may lack a mobile phone number in the prompt word template received by the large model. At this time, the large model judges that the reply lacks a key parameter, and then initiates an inquiry about the mobile phone number, and after obtaining the mobile phone number, initiates an external system call to query the customer's credit limit according to the mobile phone number, and organizes a language reply to the customer.

[0097] In another embodiment, after the large model obtains the query information in step 7 and synthesizes the existing information, if it is judged that the query information does not correspond to the existing information, it is determined to be an abnormal malicious consultation.

[0098] Example 8: {"question": "What is the loan interest rate?", "phone": null, "channel": "WeChat", "name": "Zhang San"}

[0099] To the large model: "The user is consulting the question: What is the loan interest rate? Please act as a loan consultation manager of XX Bank and briefly answer the user's question"

[0100] The large model replies: "You can query on our APP, the specific operation is XXXX"

[0101] The user continues to ask: "I am not convenient to operate now, can you help me query?"

[0102] The large model's reply: "Yes, but I haven't received your mobile phone number, would you please provide it?"

[0103] User answer: "Yes, my mobile number is: 135xxxxx"

[0104] Large model replies to special symbols for our system: "<query_credit>, arg: [phone = 135xxxx]"

[0105] After the system query, it is found that the mobile number is not bound to the user name, and the string "It is detected that the mobile number is not owned by Mr. Zhang San" is sent.

[0106] After the large model synthesizes the information given by our system, it replies: "Hello, are you Mr. Zhang San? The mobile number you provided is not Mr. Zhang San's own mobile number. Please check if the number is wrong."

[0107] User answer: "OK, no need"

[0108] As can be seen, the present embodiment can identify some malicious information extraction situations through model analysis.

[0109] If the customer is in the visitor mode, the prompt word template received by the large model not only lacks the mobile number but also lacks the name and other key information, then the large model will inquire the key information one by one, and call the external system to obtain the information to verify whether the customer's reply is corresponding, if corresponding, then answer the question according to the customer's demand, if not corresponding, then ask the customer whether the information is wrong, if the customer's reply is not corresponding for several times, then judge that the customer has no query right, after sending the ending language, stop the dialogue.

[0110] In another embodiment, a bank intelligent customer service question and answer system based on a large model is involved, as shown in Figure 3 , including a client, a system end, a vector database and a large model port, the client is used for the customer to interact with the system end, the system end is used to execute a bank intelligent customer service question and answer method based on a large model, the vector database is used to save the high-dimensional vector of the existing question and the standard answer, and the large model port is used to receive the system end instruction and call the corresponding large model to reply to the question and return the result.

[0111] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical scheme concept of the present application, some modifications and improvements can be made, which are within the protection scope of the present application.

Claims

1. A bank intelligent customer service question-and-answer method based on a large model, characterized in that, include: Step 1: Determine whether the customer's inquiry is a regulatory red line issue. If so, return the standard answer directly; otherwise, proceed to Step 2. Step 2: Route the question to the specified prompt word template based on the keywords in the question; Step 3: Process the received data to obtain the input parameters of the prompt words required for the prompt word template; Step 4: Assemble the prompt words into the specified prompt word template to obtain the complete utterance, send it to the large model, and obtain the result returned by the large model; Step 5: If the result returned by the large model contains a specific symbol, then query the corresponding information from the user database based on the content following the specific symbol and return it to the large model. The large model will then consider and respond again. If the specific symbol does not exist, then return the result returned by the large model directly to the customer. Step 6: Repeat steps 1 to 5 until the user actively ends the inquiry.

2. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, Step 1 includes: Step 1.1: Receive the consultation request sent by the user through the client, which includes the question string sent by the user and the user's basic information; Step 1.2: Extract the user's question string and convert it into a high-dimensional vector using a text-to-vector model; Step 1.3: Retrieve the high-dimensional vector obtained in Step 1.2 from the vector database, select the closest question, and determine whether the cosine similarity between the two is higher than the threshold. If it is higher, return the standard answer directly; otherwise, proceed to Step 2. The vector database stores the regulatory red line questions.

3. The intelligent customer service question-and-answer method for banks based on a large model according to claim 2, characterized in that, Step 1.2 includes: Step 1.2.1: Obtain the user's question string; Step 1.2.2: Remove redundant spaces and special characters that do not affect the semantics; Step 1.2.3: The processed string is fed into the text-to-vector model to obtain a high-dimensional vector.

4. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, The prompt word template also includes large model parameters, including large model type, large model temperature, system, and maximum token.

5. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, Step 3 includes: processing the received data to obtain the input parameters of the prompt word template; processing the original parameters sent by the upstream system to obtain the processed parameters. The processing includes structuring the data and performing numerical calculations, and determining the input parameters of the prompt word from the processed parameters according to the needs of the prompt word template.

6. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, Steps 2 and 3 are implemented using Groovy scripts. If the Groovy scripts are changed during execution, they are compiled first and then loaded into memory.

7. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, Step 4 includes: Step 4.1: The first dialogue generates an opening message for the large model based on the prompt word template; Step 4.2: Subsequent dialogues directly send user questions to the large model.

8. The intelligent customer service question-and-answer method for banks based on a large model according to claim 1, characterized in that, If the customer is in guest mode and the prompt template received by the big model is missing key input parameters such as phone number and name, the big model will query the key information one by one and call the external system to obtain information to verify whether the customer's reply matches. If they match, the big model will answer the questions according to the customer's needs. If they do not match, the big model will ask the customer if the information is incorrect. If the customer's reply does not match after multiple replies, it will be determined that the customer does not have query permission and the dialogue will stop after issuing a closing statement.

9. A bank intelligent customer service question-and-answer system based on a large model, comprising: The system comprises a client, a system, a vector database, and a large model port. The client is used for interaction between the client and the system. The system is used to execute the bank intelligent customer service question-and-answer method based on a large model as described in any one of claims 1-8. The vector database is used to store high-dimensional vectors of existing questions and standard answers. The large model port is used to receive instructions from the system and call the corresponding large model to answer the questions and return the results.

Images