Credit business system evolution method and device based on large model RAG technology

Abstract

The invention discloses a credit business system evolution method and device based on a large model RAG technology, and the method comprises the following steps: obtaining information resources, generating a theme knowledge base, generating a front knowledge base, and storing the information resources in a vector database; obtaining a question text, judging whether a reply text of the question vector can be retrieved in the front knowledge base, and if yes, obtaining the reply text; if not, retrieval is carried out through the front knowledge base and the routing theme knowledge base, and a standard text is obtained; inputting the reply text or the standard text into the large model, and obtaining a result text based on a preset output template; and based on the output result text, judging whether the reply text exists in the preposed knowledge base, if not, updating the preposed knowledge base, and if yes, updating the preposed knowledge base for the second time. In the evolution process, the product design efficiency is improved, the research and development quality is enhanced, the system operation and maintenance timeliness is improved, unnecessary human input is reduced, and the purposes of reducing cost and improving efficiency are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of financial technology, and specifically relates to a credit business system evolution method and device based on large model RAG technology. Background Art

[0002] In the contemporary banking sector, online credit business plays an extremely important role. It is not only one of the core businesses of banks, but also a key way for banks to achieve profitability, serve the real economy, and promote the flow of social funds. In order to meet the growing and changing needs of various funds, and under the requirements of supervision, banks' personal and corporate credit products continue to develop and innovate, and continue to launch financial products of different forms, types, and sizes.

[0003] In order to launch new credit products in a timely manner, it is necessary to ensure the timeliness of product development, which requires rapid deconstruction of the business logic of new products, comparison with existing product businesses, distinction between common business logic and special business logic, and accurate linkage of business modules and software components at the design and implementation level of system evolution, and rapid identification, assembly and expansion of reusable system functional components. In addition, the reliability of the system and the timeliness of operation and maintenance need to be considered, which requires timely discovery of anomalies, accurate problem location, and rapid iterative repair of the normal operation of new products and existing businesses after the new products are launched.

[0004] The credit business system needs to support online business processes and batch business processes for multi-channel and multi-type credit products. Different products have interface heterogeneity of upstream and downstream systems, process heterogeneity between functional modules, and logical heterogeneity within functional modules. At the same time, there is partial business overlap between products. According to the software architecture requirements of the product, it is necessary to achieve common interface reuse, module reuse, and logic code reuse. Based on the above characteristics of the credit business and its system, when developing new products, it is necessary to analyze existing online credit products from the business perspective, system perspective, and data perspective, and then carry out product demand design and system evolution design to ensure the feasibility of the product implementation plan, the accuracy of resource cost investment, and the impact of changes on the original functions.

[0005] In the ever-evolving and changing field of credit business systems, if we directly rely on general pre-trained large language models and their technologies to support R&D and operation and maintenance work, we will have the following limitations:

[0006] 1. Limitations of general models: General pre-trained large language models usually lack the private knowledge of specific industries and enterprises. In areas such as banking credit, these models cannot directly solve the specific credit business and system-related problems of each bank, and additional methods are needed to enable the model to learn this private knowledge.

[0007] 2. Fine-tuning technology is expensive: Fine-tuning technology is to incrementally train the large language model itself, adjust and solidify the model parameters, and enable the model to learn private knowledge. For the evolving credit products and systems, whenever there are new additions or changes to the business, using fine-tuning technology to train the model will bring a large amount of time and computing cost investment, which limits the widespread application and frequent updates of fine-tuning technology.

[0008] 3. Shortcomings of prompt word engineering: Unlike fine-tuning technology, prompt word engineering does not require expensive model training. It relies on the user to design and add language structure and background information when asking questions to prompt and control the rules and forms of the large language model output. However, the effectiveness of prompt word engineering depends largely on the design of prompt words, which requires users to have an in-depth understanding of the model and domain knowledge. In addition, this method also has shortcomings such as data bias and knowledge limitations.

[0009] 4. Advantages and dilemmas of RAG technology: RAG retrieval enhancement technology is to obtain private domain knowledge by searching the proprietary knowledge base before calling the large language model, so that the large model can reason in combination with the knowledge context. Compared with fine-tuning and prompt word technology, this method is low-cost and can effectively alleviate the large model hallucination (LLM Hallucination) and increase the accuracy and professionalism of reasoning. However, retrieval methods and architectures that have not been effectively designed may lead to defects such as poor timeliness of knowledge retrieval, low completeness of knowledge aggregation, and insufficient accuracy of knowledge association due to improper selection of document segmentation methods, text embedding methods, and retrieval matching algorithms, as well as insufficient design of knowledge organization methods and knowledge retrieval processes. In addition, due to the lack of an automatic feedback update mechanism, users cannot directly correct answer errors and participate in the adjustment of the knowledge base, which affects the timely update of knowledge and the long-term accuracy of the model. Summary of the invention

[0010] The present invention provides a credit business system evolution method and device based on large model RAG technology to solve the problems described in the background technology.

[0011] To achieve the above object, the present invention provides the following technical solution: a credit business system evolution method and device based on large model RAG technology, comprising the following steps:

[0012] Acquire information resources, classify the information resources, obtain multiple groups of subject resources, synchronously input multiple groups of subject resources and the acquired manually input question and answer corpus into the big model, output multiple subject texts through the big model, generate a subject knowledge base, refine the question and answer corpus, generate a pre-knowledge base, associate the subject knowledge base with the pre-knowledge base, and store them in the vector database;

[0013] Get the question text, extract the precise words from the question text, and vectorize the precise words to get the question vector. Determine whether the answer text of the question vector can be retrieved in the pre-knowledge base. If so, get the answer text. If not, retrieve through the pre-knowledge base and the routing subject knowledge base to get the standard text. Input the answer text or the standard text into the big model and get the result text based on the preset output template.

[0014] Based on the output result text, the manually input feedback information is obtained, and the previous knowledge base is called to determine whether the reply text exists in the previous knowledge base. If not, the question vector, result text and feedback information are updated to the previous knowledge base. If it exists, the previous knowledge base is updated for the second time.

[0015] Preferably, the outputting of multiple subject texts comprises the following steps:

[0016] Obtain question-answering corpus, obtain question-answering training text and question-answering results, where the question-answering results include positive results and negative results;

[0017] Obtain positive results and corresponding question-answering training texts, obtain question-answering training texts, vectorize the question-answering training texts, and obtain refined knowledge corpus;

[0018] Vectorize and refine knowledge corpus to obtain hot knowledge texts;

[0019] According to the hot knowledge text, obtain the matching keywords and topic tags in the question and answer corpus;

[0020] Vectorize keywords and topic tags to obtain keyword vectors and tag vectors.

[0021] Preferably, associating the subject knowledge base with the pre-knowledge base comprises the following steps:

[0022] Obtain multiple groups of subject texts, vectorize multiple subject texts respectively, generate a subject knowledge base, obtain hot knowledge texts, and build internal associations between hot knowledge texts and the subject knowledge base;

[0023] Obtain keywords and topic tags, associate keyword vectors and tag vectors with hot knowledge texts, generate a pre-knowledge base, establish a hot set, and store the hot knowledge texts in the hot set in the pre-knowledge base;

[0024] Establish a topic set in the pre-knowledge base, and store the keyword vector and label vector in the topic set;

[0025] The subject knowledge base and the prior knowledge base are stored in the vector database.

[0026] Preferably, the step of determining whether the reply text can be retrieved from the pre-requisite knowledge base comprises the following steps:

[0027] Get the question vector;

[0028] Match the question vector in the hot knowledge text of the front database, and determine whether the retrieved reply text exists in the front database according to the preset similarity threshold. If so, return the reply text and obtain the subject label of the subject knowledge base corresponding to the reply text;

[0029] If it does not exist, call the question knowledge base for retrieval.

[0030] Preferably, the retrieval through the pre-knowledge base routing subject knowledge base includes the following steps:

[0031] Get precise vocabulary;

[0032] Retrieve the topic set based on the precise vocabulary, obtain the corresponding keyword vector, obtain the topic label based on the keyword vector, and obtain the target topic knowledge base based on the topic label;

[0033] Get the question vector;

[0034] Retrieve the target subject knowledge base based on the question vector, and obtain the knowledge text segment whose similarity reaches the set threshold according to the preset similarity threshold;

[0035] Obtain the subject text corresponding to the knowledge text segment to obtain the standard text.

[0036] Preferably, the step of inputting the reply text or the standard text into the macro model to obtain the result text comprises:

[0037] To input the reply text into the large model and obtain the result text, the following steps need to be performed:

[0038] Obtain the question text, obtain the topic keywords in the question text, match the keywords according to the topic keywords, and obtain the corresponding topic tags;

[0039] Combined with the topic tags, the manually input output template is obtained, the reply text and the prompt words are synchronously input into the large model, and the result text is output through the output template;

[0040] To input the standard text into the large model and obtain the result text, the following steps need to be performed:

[0041] Obtain the question text, obtain the topic keywords in the question text, match the keywords according to the topic keywords, and obtain the corresponding topic tags;

[0042] Combined with the topic tags, the manually input output template is obtained to obtain the knowledge text segment, the knowledge text segment and the prompt word are synchronously input into the large model, and the result text is output through the output template.

[0043] Preferably, the following operations need to be performed before searching the pre-requisite knowledge base to see whether the question text exists:

[0044] Obtain feedback information and determine whether the feedback information is positive feedback or negative feedback;

[0045] If it is positive feedback, search the previous knowledge base to see if the question text exists;

[0046] If it is negative feedback, error correction information is obtained.

[0047] Preferably, the second updating of the pre-set knowledge base comprises the following steps:

[0048] Obtain feedback information and determine whether the feedback information is positive feedback or negative feedback;

[0049] If it is positive feedback, the question text, question vector and result text are obtained and updated to the previous knowledge base;

[0050] If it is negative feedback, perform the following steps:

[0051] Delete the result text in the previous knowledge base;

[0052] Obtain error correction information, vectorize the error correction information, and obtain an error correction vector;

[0053] Update the question text and error correction vector to the prior knowledge base.

[0054] A credit business system evolution device based on large model RAG technology, comprising:

[0055] The resource preprocessing module is used to obtain information resources, classify the information resources, obtain multiple groups of subject resources, synchronously input the multiple groups of subject resources and the obtained manually input question and answer corpus into the large model, output multiple subject texts, generate a subject knowledge base, refine the question and answer corpus, generate a pre-knowledge base, associate the subject knowledge base with the pre-knowledge base, and store it in the vector database;

[0056] An input processing module is used to extract precise words from the question text and vectorize the precise words to obtain a question vector;

[0057] The data retrieval module is used to determine whether the answer text of the question vector can be retrieved in the pre-knowledge base. If so, the answer text is obtained; if not, the pre-knowledge base is used to route the subject knowledge base for retrieval to obtain the standard text;

[0058] The result output module is used to input the reply text or standard text into the large model and obtain the result text based on the preset output template;

[0059] The feedback verification module is used to obtain the manually input feedback information based on the output result text, call the pre-set knowledge base, and search the pre-set knowledge base;

[0060] The feedback processing module is used to determine whether there is a retrieval result of the reply text in the previous knowledge base. If not, the question vector, result text and feedback information are updated to the previous knowledge base. If so, the previous knowledge base is updated for the second time.

[0061] A computing device includes a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the processor implements a credit business system evolution method based on large model RAG technology when executing the computer program.

[0062] The technical effects and advantages of the present invention are as follows: through the autonomous learning and manual intervention of the big model, the ability of the big model to learn and understand business domain knowledge is improved, and the R&D efficiency, quality and operation and maintenance timeliness of iterative evolution are improved; by giving priority to the front knowledge base and linking the subject knowledge base through the hot knowledge text, the result text is obtained quickly and accurately, and the feedback efficiency is improved; after outputting the result text, manual feedback information is obtained, and the hot knowledge text is continuously improved according to the feedback information, and the front knowledge base is improved to ensure that the front knowledge base can be updated in time, and the accuracy and completeness of the output content of the big model are guaranteed; in the evolution process, product design efficiency is improved, R&D quality is enhanced, system operation and maintenance timeliness is improved, and unnecessary manpower input is reduced, so as to achieve the purpose of reducing costs and increasing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the stage method flow of the present invention;

[0064] Figure 2 It is a schematic diagram of the overall method execution flow of the present invention;

[0065] Figure 3 A schematic diagram of the process of constructing a private domain knowledge base of the present invention;

[0066] Figure 4 A schematic diagram of the flow chart of the large model retrieval reasoning method of the present invention;

[0067] Figure 5 It is a schematic diagram of the knowledge base feedback tuning method flow of the present invention;

[0068] Figure 6 It is a system schematic diagram of the present invention. DETAILED DESCRIPTION

[0069] 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.

[0070] The present invention provides Figure 1-2 The method and device for credit business system evolution based on large model RAG technology shown in the embodiment of the present invention perform the above steps through the following modules: resource preprocessing module, input processing module, data retrieval module, feedback processing module, feedback verification module and feedback processing module. The method of the present invention is performed by using the above modules, including a private domain knowledge base construction stage, a large model retrieval reasoning stage and a knowledge base feedback tuning stage. The specific execution includes the following steps:

[0071] S1: In the construction phase of the private domain knowledge base, information resources are obtained through the resource preprocessing module, where the information resources are content related to the credit business system, including: business documents, design documents, operation and maintenance records, code files, etc. Information resources are input and the private domain information resources are classified and summarized, that is: business documents, design documents, operation and maintenance records, code files and other resources are classified and summarized according to business themes, R&D themes, and operation and maintenance themes to obtain multiple groups of subject resources and store them in the file server; multiple subject resources are associated and integrated from different perspectives such as credit business lines and credit product lines, and imported into the general large model in the form of attachments, so that the large model can learn and understand autonomously, and obtain question and answer corpus through manual input, including; keywords, output templates and question and answer texts, etc., conduct multiple rounds of large model question and answer on the large model, and guide the large model to summarize and refine the correct understanding of the subject resources through feedback of positive and negative results to obtain the subject text;

[0072] S11: Obtain Q&A corpora, and extract Q&A training texts and Q&A results. The Q&A results include positive results and negative results, which are used as refined knowledge corpora. Among them, the topic knowledge base is divided into a product topic knowledge base, a R & D topic knowledge base, and an operation and maintenance topic knowledge base. Obtain the refined knowledge corpus, vectorize the refined knowledge corpus to obtain hot knowledge texts, store the hot knowledge texts in the topic knowledge base, design the metadata structure to build the internal association between the hot knowledge texts and the topic knowledge base, and establish the external association between the refined knowledge corpus and the topic texts. Among them, in the metadata structure design, the association number between the hot knowledge texts and the topic texts can be set to quickly query the corresponding content of the hot knowledge texts in the topic texts during subsequent execution. Obtain Q&A corpora, obtain keywords and topic tags, and vectorize them to obtain keyword vectors and tag vectors, and associate the keyword vectors and tag vectors with the hot knowledge texts respectively to generate a pre-knowledge base.

[0073] S12: Obtain multiple topic texts, vectorize the multiple topic texts to generate a topic knowledge base. Associate the keyword vectors and tag vectors with the hot knowledge texts to generate a pre-knowledge base. Establish a hot set in the pre-knowledge base, store the hot knowledge texts in the hot set of the pre-knowledge base, and establish a topic set in the pre-knowledge base, and store the keyword vectors and tag vectors in the topic set.

[0074] S13: Store the topic knowledge base and the pre-knowledge base in a vector database.

[0075] In this way, the pre-knowledge base associates hot knowledge texts based on keyword vectors and tag vectors, can lock the pre-knowledge base in the vector database, and quickly retrieve the answers to hot questions. If there is no answer in the pre-knowledge base, in-depth retrieval can be performed through the topic knowledge base to obtain the answer. Moreover, the pre-knowledge base also provides the storage ability of hot knowledge texts. The hot knowledge texts include credit hot private domain knowledge, feedback information knowledge, etc., and associate and store the hot knowledge texts in the hot set of the pre-knowledge base.

[0076] S2: In the large model retrieval and reasoning stage, obtain the question text through the input processing module, perform word segmentation on the question text, exclude common words such as "of, is, right", etc., to obtain precise words, vectorize the precise words to obtain a question vector, and retrieve the question vector in the pre-database through the data retrieval module. According to the preset threshold, judge whether a reply text can be retrieved from the hot knowledge texts in the pre-database. For example, using the cosine similarity algorithm, the preset threshold cos(θ)≥0.8, including the following two situations:

[0077] S21: When a reply text whose similarity score reaches a set threshold is retrieved, that is, when the score cos(θ)≥0.8, a reply text exists, and step S23 is executed;

[0078] S22: When the similarity score fails to reach the set threshold, i.e., cos(θ) < 0.8, or when no hot knowledge is hit, there is no reply text, and the search will be performed through the subject knowledge base routing subject knowledge base, and the following operations need to be performed:

[0079] S221: Obtain precise words, search the subject set of the pre-set knowledge base according to the precise words, and obtain the subject tags corresponding to the precise words. For example, if the input is: "What is the product code of loan assistance product A? Please locate the special code branch of product A in the software?", the precise words parsed include "loan assistance product A", "product code", and "code branch". The subject set is searched and the keyword vectors such as 'code branch' and 'product code' are matched, and then the subject tags of the product knowledge base and the R&D knowledge base are obtained from the corresponding metadata;

[0080] S222: Retrieve the target subject knowledge base based on the question vector, and according to the preset similarity threshold, obtain the knowledge text segment whose similarity reaches the set threshold, and obtain the subject text corresponding to the knowledge text segment to obtain the standard text; wherein, the refined knowledge corpus includes question text and answer text, which are stored separately and segmented, therefore, one or more knowledge text segments may be hit during the retrieval process, and the metadata of the knowledge text segments are obtained respectively, and the associated number of the corresponding knowledge text segment is obtained, and combined with the filtering query function of the vector knowledge base, the subject text associated with all the knowledge text segments is obtained as the standard text.

[0081] S23: Obtain the reply text or standard text through the result output module. If it is the reply text retrieved from the previous knowledge base, obtain the question text, obtain the subject keywords in the question text, match the keywords according to the subject keywords, obtain the corresponding subject tags, and obtain the manually input output template based on the subject tags to obtain prompt words, and input the reply text and the prompt words into the big model synchronously, and output the result text based on the output template; if it is the standard text retrieved from the subject knowledge base, obtain the question text, obtain the subject keywords in the question text, match the keywords according to the subject keywords, obtain the corresponding subject tags, and input the knowledge text segment and the prompt words into the big model synchronously with the subject tags. After the big model receives the input content, it will combine the subject tags to obtain the manually input output template to obtain prompt words, and use the contents of the knowledge text segment and the prompt words to improve the output content of the big model, and output the result text based on the output template. When the big model outputs the result text, the topic text is called as the big model reasoning corpus to avoid using it directly for big model reasoning, which may cause the big model to "take it out of context", produce model hallucinations, and output inaccurate situations, so as to ensure that the content input to the big model is accurate and comprehensive, and then guide the model to generate accurate and user-friendly answers through the output template, and output the result text. Take the output template for building the R&D theme as an example, such as: "Based on the following context: {Context}, answer the R&D theme question: {Question}, output the answer in the example format, and list the theme text based on it. The example is: the name of the product is: X; the software components and classes involved include: P1, P2; the theme text based on it includes: S1, S2".

[0082] S3: In the knowledge base feedback tuning stage, after obtaining the result text, the feedback processing module obtains the feedback information input by the business line, R&D line, and other human-computer interaction methods, and vectorizes the feedback information to obtain the feedback vector. After that, the feedback vector is verified by the feedback verification module, and the following two situations are obtained:

[0083] S31: negative feedback, then manually input error correction information is obtained;

[0084] S32: Positive feedback indicates that the output standard text meets the user's needs. The question text, question vector and standard text are extracted, and based on the vector similarity algorithm, the feedback processing module is called to search for the reply text in the pre-knowledge base to determine whether the result with a matching degree reaching the set threshold is retrieved. If not, the question text, standard text and feedback information are directly stored in the hot knowledge text;

[0085] If the search results in text, execute the following steps:

[0086] S311: Determine whether the feedback information is positive feedback or negative feedback. If it is positive feedback, execute step S312; if it is negative feedback, execute step S313;

[0087] S312: If it is positive feedback, the question text, question vector and result text are obtained, and this conversation will be used as one of the reference answers to the current question, and the question text, question vector and result text are updated to the hot knowledge text of the pre-set knowledge base, wherein the result text is the reply text or the standard text;

[0088] S313: If it is negative feedback, then obtain the hot knowledge text, delete the corresponding question text and answer text that already exist in the hot knowledge text in the pre-positioned knowledge base, and update the current question text and feedback information to the pre-positioned knowledge base based on the result text, and update the hot knowledge text. In this way, the knowledge base is continuously optimized, thereby improving the recall rate of retrieval, enhancing the performance of retrieval, and improving the accuracy of knowledge. By timely updating the question text and feedback information, the knowledge base is continuously optimized and updated, and the accuracy of the output result text is continuously enhanced.

[0089] Among them, the similarity algorithm can be optimized in the algorithm set supported by the vector database, such as inner product, Euclidean distance, cosine similarity and other algorithms.

[0090] To sum up, through the above methods, the R&D efficiency, quality, and operation and maintenance timeliness of the iterative evolution of the credit system can be improved. Moreover, during the evolution of the credit system, the product design efficiency can be improved, the R&D quality can be enhanced, the system operation and maintenance timeliness can be improved, and unnecessary manpower investment can be reduced, thereby achieving the goal of reducing costs and increasing efficiency.

[0091] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A credit business system evolution method based on large model RAG technology, characterized by: The steps include: Acquire information resources, classify the information resources, obtain multiple groups of subject resources, synchronously input multiple groups of subject resources and the acquired manually input question and answer corpus into the big model, output multiple subject texts through the big model, generate a subject knowledge base, refine the question and answer corpus, generate a pre-knowledge base, associate the subject knowledge base with the pre-knowledge base, and store them in the vector database; Get the question text, extract the precise words from the question text, and vectorize the precise words to get the question vector. Determine whether the answer text of the question vector can be retrieved in the pre-knowledge base. If so, get the answer text. If not, retrieve through the pre-knowledge base and the routing subject knowledge base to get the standard text. Input the answer text or the standard text into the big model and get the result text based on the preset output template. Based on the output result text, the manually input feedback information is obtained, and the previous knowledge base is called to determine whether the reply text exists in the previous knowledge base. If not, the question vector, result text and feedback information are updated to the previous knowledge base. If it exists, the previous knowledge base is updated for the second time.

2. The credit business system evolution method based on large model RAG technology according to claim 1 is characterized by: The output of multiple subject texts comprises the following steps: Obtain question-answering corpus, obtain question-answering training text and question-answering results, where the question-answering results include positive results and negative results; Obtain positive results and corresponding question-answering training texts, obtain question-answering training texts, vectorize the question-answering training texts, and obtain refined knowledge corpus; Vectorize and refine knowledge corpus to obtain hot knowledge texts; According to the hot knowledge text, obtain the matching keywords and topic tags in the question and answer corpus; Vectorize keywords and topic tags to obtain keyword vectors and tag vectors.

3. The credit business system evolution method based on large model RAG technology according to claim 2 is characterized by: The step of associating the subject knowledge base with the pre-knowledge base comprises the following steps: Obtain multiple groups of subject texts, vectorize multiple subject texts respectively, generate a subject knowledge base, obtain hot knowledge texts, and build internal associations between hot knowledge texts and the subject knowledge base; Obtain keywords and topic tags, associate keyword vectors and tag vectors with hot knowledge texts, generate a pre-knowledge base, establish a hot set, and store the hot knowledge texts in the hot set in the pre-knowledge base; Establish a topic set in the pre-knowledge base, and store the keyword vector and label vector in the topic set; The subject knowledge base and the prior knowledge base are stored in the vector database.

4. The credit business system evolution method based on large model RAG technology according to claim 3 is characterized by: The step of judging whether the reply text can be retrieved from the pre-requisite knowledge base comprises the following steps: Get the question vector; Match the question vector in the hot knowledge text of the front database, and determine whether the retrieved reply text exists in the front database according to the preset similarity threshold. If so, return the reply text and obtain the subject label of the subject knowledge base corresponding to the reply text; If it does not exist, call the question knowledge base for retrieval.

5. The credit business system evolution method based on large model RAG technology according to claim 4 is characterized by: The method of searching through the pre-knowledge base routing the subject knowledge base includes the following steps: Get precise vocabulary; Retrieve the topic set based on the precise vocabulary, obtain the corresponding keyword vector, obtain the topic label based on the keyword vector, and obtain the target topic knowledge base based on the topic label; Get the question vector; Retrieve the target subject knowledge base based on the question vector, and obtain the knowledge text segment whose similarity reaches the set threshold according to the preset similarity threshold; Obtain the subject text corresponding to the knowledge text segment to obtain the standard text.

6. The credit business system evolution method based on large model RAG technology according to claim 5 is characterized by: The step of inputting the reply text or standard text into the large model to obtain the result text includes: To input the reply text into the large model and obtain the result text, the following steps need to be performed: Obtain the question text, obtain the topic keywords in the question text, match the keywords according to the topic keywords, and obtain the corresponding topic tags; Combined with the topic tags, the manually input output template is obtained, the reply text and the prompt words are synchronously input into the large model, and the result text is output through the output template; To input the standard text into the large model and obtain the result text, the following steps need to be performed: Obtain the question text, obtain the topic keywords in the question text, match the keywords according to the topic keywords, and obtain the corresponding topic tags; Combined with the topic tags, the manually input output template is obtained to obtain the knowledge text segment, the knowledge text segment and the prompt word are synchronously input into the large model, and the result text is output through the output template.

7. The credit business system evolution method based on large model RAG technology according to claim 6 is characterized by: Before searching whether the question text exists in the pre-requisite knowledge base, the following operations need to be performed: Obtain feedback information and determine whether the feedback information is positive feedback or negative feedback; If it is positive feedback, search the previous knowledge base to see if the question text exists; If it is negative feedback, error correction information is obtained.

8. The credit business system evolution method based on large model RAG technology according to claim 7 is characterized by: The second updating of the pre-set knowledge base comprises the following steps: Obtain feedback information and determine whether the feedback information is positive feedback or negative feedback; If it is positive feedback, the question text, question vector and result text are obtained and updated to the previous knowledge base; If it is negative feedback, perform the following steps: Delete the result text in the previous knowledge base; Obtain error correction information, vectorize the error correction information, and obtain an error correction vector; Update the question text and error correction vector to the prior knowledge base.

9. A credit business system evolution device based on large model RAG technology, characterized by: include: The resource preprocessing module is used to obtain information resources, classify the information resources, obtain multiple groups of subject resources, synchronously input the multiple groups of subject resources and the obtained manually input question and answer corpus into the large model, output multiple subject texts, generate a subject knowledge base, refine the question and answer corpus, generate a pre-knowledge base, associate the subject knowledge base with the pre-knowledge base, and store it in the vector database; An input processing module is used to extract precise words from the question text and vectorize the precise words to obtain a question vector; The data retrieval module is used to determine whether the answer text of the question vector can be retrieved in the pre-knowledge base. If so, the answer text is obtained; if not, the pre-knowledge base is used to route the subject knowledge base for retrieval to obtain the standard text; The result output module is used to input the reply text or standard text into the large model and obtain the result text based on the preset output template; The feedback verification module is used to obtain the manually input feedback information based on the output result text, call the pre-set knowledge base, and search the pre-set knowledge base; The feedback processing module is used to determine whether there is a retrieval result of the reply text in the previous knowledge base. If not, the question vector, result text and feedback information are updated to the previous knowledge base. If so, the previous knowledge base is updated for the second time.

10. A computing device comprising a memory and a processor, characterized in that: The memory stores executable codes, and when the processor executes the executable codes, the processor implements the methods described in claims 1-8 when executing the computer program.

Images