A credit service system evolution method and device based on a large model RAG technology

Abstract

The application discloses a credit service system evolution method and device based on a large model RAG technology, comprising the following steps: acquiring information resources, generating a theme knowledge base and a pre-knowledge base, and storing them to a vector database; acquiring a question text, judging whether a reply text of the question vector can be searched in the pre-knowledge base, if yes, the reply text is obtained; if no, the pre-knowledge base is used to route the theme knowledge base for searching, and a standard text is obtained; the reply text or the standard text is input into a large model, a result text is obtained based on a preset output template; whether the reply text exists in the pre-knowledge base is judged based on the output result text, if not, the pre-knowledge base is updated, if yes, the pre-knowledge base is updated again. In the evolution process, the product design efficiency is improved, the research and development quality is enhanced, the system operation efficiency is improved, the unnecessary human input is reduced, and the cost reduction and benefit increase purposes are achieved.

Description

Technical Field

[0001] This invention belongs to the field of financial technology, specifically relating to a method and apparatus for the evolution of a credit business system based on large model RAG technology. Background Technology

[0002] In the contemporary banking sector, online lending plays a vital role. It is not only one of the core businesses of banks but also a key channel for them to achieve profitability, serve the real economy, and promote the flow of social capital. To meet the ever-growing and ever-changing funding needs of various entities, and in compliance with regulatory requirements, banks are continuously developing and innovating their personal and corporate lending products, launching a constant stream of financial products in different forms, types, and sizes.

[0003] To ensure the timely launch of new credit products, it is essential to guarantee the timeliness of product development. This requires rapid deconstruction of the new product's business logic, comparison with existing product business logic, differentiation of common and specific business logic, and accurate linking of business modules and software components at the system evolution design and implementation levels. It also necessitates the rapid identification, assembly, and expansion of reusable system functional components. Furthermore, system reliability and timely operation and maintenance must be considered. This requires timely anomaly detection, accurate problem localization, and rapid iterative repair after the new product's launch, ensuring the normal operation of both the new product and existing business.

[0004] The credit business system needs to support online and batch business processes for multiple channels and types of credit products. Different products exhibit heterogeneity in interfaces between upstream and downstream systems, process heterogeneity between functional modules, and logical heterogeneity within functional modules. Simultaneously, there is partial overlap in business processes among the products. Based on the product's software architecture requirements, common interface reuse, module reuse, and logic code reuse are necessary. Given these characteristics of the credit business and its system, developing new products requires analyzing existing online credit products from business, system, and data perspectives. Following this, product requirement design and system evolution design should be conducted to ensure the feasibility of the product implementation plan, the accuracy of resource cost allocation, and the impact of changes on existing functions.

[0005] In the ever-evolving field of credit business systems, directly relying on general-purpose pre-trained large language models and their technologies to support research and development and operations will have the following limitations:

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

[0007] 2. High Cost of Fine-Tuning Techniques: Fine-tuning techniques involve incrementally training the large language model itself, adjusting and solidifying model parameters to enable the model to acquire private knowledge. For constantly evolving credit products and systems, whenever there are new additions or changes in business, training the model using fine-tuning techniques incurs significant time and computational costs, which limits the widespread application and frequent updates of fine-tuning techniques.

[0008] 3. Limitations of Cue Word Engineering: Unlike fine-tuning techniques, cue word engineering does not require expensive model training. It relies on users designing and adding language structure and background information when asking questions to prompt and control the rules and form of the large language model's output. However, the effectiveness of cue word engineering largely depends on the design of the cue words, which requires users to understand the model and have a deep grasp of domain knowledge. In addition, this method also suffers from limitations such as data bias and knowledge limitations.

[0009] 4. Advantages and Challenges of RAG Technology: RAG retrieval enhancement technology retrieves private domain knowledge from a proprietary knowledge base before calling the large language model, allowing the large model to reason within the knowledge context. Compared to fine-tuning and prompt word techniques, this method is lower in cost and can effectively alleviate the large model hallucination (LLM Hallucination), increasing reasoning accuracy and professionalism. However, poorly designed retrieval methods and architectures may suffer from poor timeliness of knowledge retrieval, low completeness of knowledge aggregation, and insufficient accuracy of knowledge association due to inappropriate document segmentation methods, text embedding methods, retrieval matching algorithms, and inadequate design of knowledge organization methods and knowledge retrieval processes. Furthermore, the lack of an automatic feedback update mechanism prevents users from directly correcting answer errors and participating in knowledge base adjustments, thus affecting timely knowledge updates and the long-term accuracy of the model. Summary of the Invention

[0010] This invention provides a method and apparatus for the evolution of a credit business system based on large model RAG technology, in order to solve the problems described in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a method and apparatus for the evolution of a credit business system based on large-scale model (RAG) technology, comprising the following steps:

[0012] Information resources are acquired, classified, and multiple sets of topic resources are obtained. These multiple sets of topic resources and the acquired question-and-answer corpus are simultaneously input into a large model. The large model outputs multiple topic texts to generate a topic knowledge base. Question-and-answer corpus is extracted to generate a pre-knowledge base. The topic knowledge base is associated with the pre-knowledge base and stored in a vector database.

[0013] The process involves obtaining the question text, extracting precise words from the question text, quantifying the precise words to obtain the question vector, determining whether the answer text of the question vector can be retrieved in the prior knowledge base, obtaining the answer text if it can, and otherwise retrieving the standard text through the prior knowledge base and the topic knowledge base. The answer text or standard text is then input into the large model, and the result text is obtained based on the preset output template.

[0014] Based on the output result text, obtain the feedback information input by the human, call the prior knowledge base, and determine whether there is a response text in the prior knowledge base. If not, update the question vector, result text and feedback information to the prior knowledge base. If they exist, update the prior knowledge base a second time.

[0015] Preferably, the output of multiple topic texts includes the following steps:

[0016] Obtain the question-and-answer corpus to obtain the question-and-answer training text and the question-and-answer results, including positive and negative results;

[0017] Obtain positive results and corresponding question-and-answer training texts to obtain question-and-answer training texts. Vectorize the question-and-answer training texts to obtain refined knowledge corpus.

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

[0019] Based on trending knowledge texts, obtain matching keywords and topic tags from the question-and-answer corpus;

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

[0021] Preferably, associating the topic knowledge base with the prior knowledge base includes the following steps:

[0022] Obtain multiple sets of topic texts, vectorize each set of topic texts to generate a topic knowledge base, obtain hot topic knowledge texts, and construct the internal relationship between hot topic knowledge texts and the topic knowledge base;

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

[0024] Establish a topic set in the prior knowledge base, and store keyword vectors and tag vectors in the topic set;

[0025] Store the subject knowledge base and the prerequisite knowledge base in a vector database.

[0026] Preferably, determining whether the response text can be retrieved in the prior knowledge base includes the following steps:

[0027] Obtain the question vector;

[0028] Match the question vector in the hot knowledge text of the pre-database, and determine whether the answer text exists in the pre-database according to the preset similarity threshold. If it exists, return the answer text and obtain the topic tags of the topic knowledge base corresponding to the answer text.

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

[0030] Preferably, the step of retrieving from the topic knowledge base via the preceding knowledge base includes the following steps:

[0031] Get precise vocabulary;

[0032] Based on the precise vocabulary search topic set, the corresponding keyword vector is obtained, the topic tags are obtained based on the keyword vector, and the target topic knowledge base is obtained based on the topic tags;

[0033] Obtain the question vector;

[0034] Based on the query vector retrieval of the target topic knowledge base, and according to the preset similarity threshold, the knowledge text segments with similarity reaching the set threshold are obtained;

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

[0036] Preferably, the step of inputting the response text or standard text into the large model to obtain the result text includes:

[0037] To input the response text into the large model and obtain the result text, the following steps are required:

[0038] Obtain the question text, extract the topic keywords from the question text, match keywords based on the topic keywords, and obtain the corresponding topic tags;

[0039] By combining topic tags, an output template for human input is obtained. The response text and prompt words are simultaneously input into the large model, and the result text is output through the output template.

[0040] To input standard text into a large model and obtain the resulting text, the following steps are required:

[0041] Obtain the question text, extract the topic keywords from the question text, match keywords based on the topic keywords, and obtain the corresponding topic tags;

[0042] By combining topic tags, an output template for manual input is obtained, and knowledge text segments are acquired. These knowledge text segments and prompt words are then simultaneously input into the large model, and the result text is output through the output template.

[0043] Preferably, before retrieving the prior knowledge base to check if the question text exists, the following steps are required:

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

[0045] If the feedback is positive, then check the prior knowledge base to see if the question text exists;

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

[0047] Preferably, the secondary update of the prior knowledge base includes the following steps:

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

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

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

[0051] Delete the result text from the prior knowledge base;

[0052] Obtain error correction information, vectorize the error correction information, and obtain the 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 acquire information resources, classify information resources, obtain multiple sets of topic resources, synchronously input multiple sets of topic resources and acquired manually input question and answer corpus into the large model, output multiple topic texts, generate topic knowledge base, extract question and answer corpus, generate pre-knowledge base, associate topic knowledge base with pre-knowledge base, and store in vector database.

[0056] The input processing module is used to extract precise words from the question text and quantize the precise words to obtain the question vector.

[0057] The data retrieval module is used to determine whether the response text of the question vector can be retrieved in the prior knowledge base. If it can, the response text is obtained; otherwise, the standard text is obtained by retrieval through the prior knowledge base and the topic knowledge base.

[0058] The results output module is used to input the response 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 feedback information input by humans based on the output result text, call the prior knowledge base, and retrieve the prior knowledge base;

[0060] The feedback processing module is used to determine whether the search results for the response text exist in the prior knowledge base. If not, the question vector, the result text, and the feedback information are updated in the prior knowledge base. If they exist, the prior knowledge base is updated a 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 executes the computer program to implement a credit business system evolution method based on large model RAG technology.

[0062] The technical effects and advantages of this invention are as follows: By combining autonomous learning with manual intervention, the large model's ability to learn and understand business domain knowledge is improved, thereby enhancing the R&D efficiency, quality, and operational timeliness of iterative evolution. By prioritizing the selection of a pre-existing knowledge base and linking it to thematic knowledge bases with hot-topic knowledge texts, the resulting text is obtained quickly and accurately, improving feedback efficiency. After outputting the resulting text, human feedback is obtained, and hot-topic knowledge texts are continuously improved based on this feedback, thus perfecting the pre-existing knowledge base and ensuring its timely updates, guaranteeing the accuracy and completeness of the large model's output. During the evolution process, product design efficiency is improved, R&D quality is enhanced, system operational timeliness is increased, and unnecessary manpower input is reduced, achieving the goal of cost reduction and efficiency improvement. Attached Figure Description

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

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

[0065] Figure 3 This is a schematic diagram of the private domain knowledge base construction method of the present invention;

[0066] Figure 4 This is a schematic diagram of the large-model retrieval and reasoning method of the present invention;

[0067] Figure 5 This is a schematic diagram of the knowledge base feedback optimization method of the present invention;

[0068] Figure 6 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] This invention provides, for example Figure 1-2 The method and apparatus for evolving a credit business system based on large model RAG technology, as shown, executes 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 this invention is executed using the aforementioned modules, including a private domain knowledge base construction stage, a large model retrieval and reasoning stage, and a knowledge base feedback optimization stage. Specifically, the execution includes the following steps:

[0071] S1: In the private domain knowledge base construction phase, information resources are acquired through the resource preprocessing module. These resources are related to the credit business system, including business documents, design documents, operation and maintenance records, and code files. The input information resources are then categorized and summarized according to business themes, development themes, and operation and maintenance themes, resulting in multiple sets of thematic resources, which are stored on a file server. Multiple thematic resources are then integrated from different perspectives, such as the credit business line and credit product line, and imported into the general large model as attachments. This allows the large model to learn and understand autonomously. Question and answer data, including keywords, output templates, and question and answer text, is obtained through manual input. Multiple rounds of large model question and answer are conducted, and feedback from positive and negative results guides the large model to summarize and refine the correct understanding of the thematic resources, resulting in thematic text.

[0072] S11: acquiring question-and-answer corpus, extracting and obtaining question-and-answer training texts and question-and-answer results, wherein the question-and-answer results include positive results and negative results, which are used as refined knowledge corpus; wherein the topic knowledge base is divided into a product topic knowledge base, a research and development topic knowledge base, and an operation and maintenance topic knowledge base; acquiring the refined knowledge corpus, vectorizing the refined knowledge corpus to obtain hot spot knowledge texts, storing the hot spot knowledge texts in the topic knowledge base, constructing internal associations between the hot spot knowledge texts and the topic knowledge base through metadata structure design, and establishing external associations between the refined knowledge corpus and topic texts, wherein in the metadata structure design, association numbers between the hot spot knowledge texts and the topic texts can be set, so that corresponding contents of the hot spot knowledge texts in the topic texts can be quickly queried in subsequent execution; acquiring the question-and-answer corpus to obtain keywords and topic tags, performing vectorization to obtain keyword vectors and tag vectors, respectively associating the keyword vectors and the tag vectors with the hot spot knowledge texts, and generating a preposed knowledge base;

[0073] S12: acquiring a plurality of topic texts, vectorizing the plurality of topic texts to generate the topic knowledge base; associating the keyword vectors and the tag vectors with the hot spot knowledge texts to generate the preposed knowledge base, establishing a hot spot set in the preposed knowledge base, storing the hot spot knowledge texts in the hot spot set of the preposed knowledge base, establishing a topic set in the preposed knowledge base, and storing the keyword vectors and the tag vectors in the topic set;

[0074] S13: storing the topic knowledge base and the preposed knowledge base in a vector database.

[0075] In this way, the preposed knowledge base associates the hot spot knowledge texts based on the keyword vectors and the tag vectors, can lock the preposed knowledge base in the vector database and quickly retrieve answers to hot spot questions. If no answer exists in the preposed knowledge base, deep retrieval can be performed through the topic knowledge base to obtain the answer. In addition, the preposed knowledge base further provides storage capacity for hot spot knowledge texts, and the hot spot knowledge texts include credit hot spot private domain knowledge, feedback information knowledge and the like, which are associated and stored in the hot spot set of the preposed knowledge base.

[0076] S2: in the large model retrieval reasoning stage, acquiring question text through an input processing module, performing word segmentation on the question text, removing general words such as "de", "shi", "ma" and the like to obtain accurate words, vectorizing the accurate words to obtain a question vector, retrieving the question vector in a preposed database through a data retrieval module, and judging whether a reply text can be retrieved from the hot spot knowledge texts of the preposed database according to a preset threshold, for example, using a cosine similarity algorithm with a preset threshold cos(θ)≥0.8, which includes the following two cases:

[0077] S21: If a response text with a similarity score that reaches the set threshold is found, that is, when the score cos(θ)≥0.8, then a response text exists and step S23 will be executed.

[0078] S22: If no similarity score is found that meets the set threshold (cos(θ) < 0.8) or if no hot topics are found, then no response text will be available. The search will be conducted through the topic knowledge base, and the following steps will be taken:

[0079] S221: Obtain precise keywords. Based on the precise keywords, search the topic set of the prior knowledge base to obtain the topic tags corresponding to the precise keywords. 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 parsed precise keywords include "loan assistance product A", "product code", and "code branch". Search the topic set and match keyword vectors such as "code branch" and "product code". Then, obtain the topic tags of the product knowledge base and R&D knowledge base from the corresponding metadata.

[0080] S222: Based on the query vector retrieval of the target topic knowledge base, according to the preset similarity threshold, acquire knowledge text segments with similarity reaching the set threshold, and acquire the corresponding topic text of the knowledge text segments to obtain the standard text; wherein, the refined knowledge corpus includes question text and answer text, which are stored separately and in segments. Therefore, one or more knowledge text segments may be hit during the retrieval process. By acquiring the metadata of each knowledge text segment, the association number of the corresponding knowledge text segment is obtained. Combined with the filtering query function of the vector knowledge base, all topic texts associated with the knowledge text segments are obtained as the standard text.

[0081] S23: Obtain the response text or standard text through the result output module. If the response text is retrieved from the prior knowledge base, obtain the question text, get the topic keywords in the question text, match keywords according to the topic keywords to obtain the corresponding topic tags, and obtain the manually input output template based on the topic tags to obtain prompt words. Input the response text and prompt words into the large model simultaneously. Output the result text based on the output template. If the response text is retrieved from the topic knowledge base, obtain the question text, get the topic keywords in the question text, match keywords according to the topic keywords to obtain the corresponding topic tags, and input the knowledge text segment and prompt words into the large model simultaneously. After receiving the input content, the large model will combine the topic tags to obtain the manually input output template to obtain prompt words. The content of the knowledge text segment and prompt words will be used to improve the output content of the large model. Output the result text based on the output template. When the large model outputs result text, it uses topic text as inference corpus to avoid directly using it for inference, which could lead to the large model "taking things out of context," creating model illusions, and resulting in inaccurate output. This ensures that the content input to the large model is accurate and comprehensive. Then, an output template guides the model to generate precise and easily understandable answers, outputting the result text. For example, to construct an output template for a research and development topic, it might be: "Based on the following context: {Context}, answer the research and development topic question: {Question}, output the answer according to the example format, and list the topic text used. Example: Product name: X; Software components and classes involved include: P1, P2; Topical text used includes: S1, S2."

[0082] S3: Knowledge base feedback optimization phase. After obtaining the result text, the feedback processing module acquires feedback information input through human-computer interaction methods such as business lines and R&D lines, and vectorizes the feedback information to obtain feedback vectors. Then, the feedback verification module verifies the feedback vectors, resulting in the following two cases:

[0083] S31: Negative feedback, which obtains error correction information input by humans;

[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 answer text in the pre-knowledge base. It is determined whether a result with a matching degree reaching the set threshold is found. If no result is found, the question text, standard text, and feedback information are directly stored in the hot knowledge text.

[0085] If the search returns text results, proceed with the following steps:

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

[0087] S312: If it is positive feedback, then obtain the question text, question vector and result text, and this session will be used as one of the reference answers to the current question. Update the question text, question vector and result text to the hot knowledge text in the prior knowledge base, where the result text is the answer text or standard text.

[0088] S313: If the feedback is negative, retrieve the hot topic knowledge text, delete the corresponding question and answer texts already existing within the hot topic knowledge text in the preceding knowledge base, and update the preceding knowledge base with the current question text and feedback information based on the result text. This updates the hot topic knowledge text, thereby continuously optimizing the knowledge base, improving retrieval recall, enhancing retrieval performance, and improving knowledge accuracy. By timely updating the question text and feedback information, the knowledge base is continuously optimized and updated, constantly enhancing the accuracy of the output result text.

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

[0090] In summary, the above methods can improve the R&D efficiency, quality, and operational timeliness of the credit system's iterative evolution. Furthermore, during the evolution of the credit system, they can enhance product design efficiency, improve R&D quality, increase system operational timeliness, and reduce unnecessary manpower input, thereby achieving the goal of cost reduction and efficiency improvement.

[0091] Finally, it should be noted that the above description 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for the evolution of a credit business system based on large-scale model (RAG) technology, characterized in that: Includes the following steps: Information resources, specifically content related to the credit business system, are acquired and categorized to obtain multiple sets of topic resources. These topic resources, along with manually inputted question-and-answer corpora, are simultaneously input into a large model. The large model outputs multiple topic texts to generate a topic knowledge base. Question-and-answer corpora are then extracted to generate a pre-existing knowledge base. The topic knowledge base is then linked to the pre-existing knowledge base, including: acquiring multiple sets of topic texts, vectorizing each set, generating the topic knowledge base; acquiring hot topic knowledge texts and establishing internal associations between these texts and the topic knowledge base; acquiring keywords and topic tags, associating keyword and tag vectors with hot topic knowledge texts to generate the pre-existing knowledge base, and establishing a hot topic set, storing the hot topic knowledge texts in the hot topic set within the pre-existing knowledge base; establishing topic sets within the pre-existing knowledge base and storing keyword and tag vectors in these topic sets; and finally, storing the topic knowledge base and the pre-existing knowledge base in a vector database. The process involves: acquiring the question text; extracting precise words from the question text and quantifying these precise words to obtain a question vector; determining whether the answer text corresponding to the question vector can be retrieved from the pre-existing knowledge base; if so, obtaining the answer text; otherwise, retrieving the standard text by routing through the topic knowledge base from the pre-existing knowledge base; inputting the answer text or standard text into the large model; and obtaining the result text based on a preset output template. The process of retrieving through the topic knowledge base from the pre-existing knowledge base includes: acquiring precise words; retrieving a topic set based on the precise words to obtain corresponding keyword vectors; obtaining topic tags based on the keyword vectors; obtaining the target topic knowledge base based on the topic tags; acquiring the question vector; retrieving the target topic knowledge base based on the question vector; acquiring knowledge text segments with similarity reaching a preset similarity threshold; and acquiring the topic text corresponding to the knowledge text segments to obtain the standard text. Based on the output result text, obtain the feedback information input by the user, call the prior knowledge base, and determine whether the response text exists in the prior knowledge base. If it does not exist, update the question vector, result text, and feedback information to the prior knowledge base. If it exists, update the prior knowledge base a second time, including: obtaining the feedback information and determining whether it is positive or negative feedback; if it is positive feedback, obtain the question text, question vector, and result text, and update them to the prior knowledge base; if it is negative feedback, perform the following steps: delete the result text from the prior knowledge base; obtain the error correction information, vectorize the error correction information to obtain the error correction vector; update the question text and error correction vector to the prior knowledge base.

2. The method for evolving a credit business system based on large-scale model (RAG) technology according to claim 1, characterized in that: The process of outputting multiple topic texts includes the following steps: Obtain the question-and-answer corpus to obtain the question-and-answer training text and the question-and-answer results, including positive and negative results; Obtain positive results and corresponding question-and-answer training texts to obtain question-and-answer training texts. Vectorize the question-and-answer training texts to obtain refined knowledge corpus. Vectorize and refine the knowledge corpus to obtain hot topic knowledge texts; Based on trending knowledge texts, obtain matching keywords and topic tags from the question-and-answer corpus; Vectorize keywords and topic tags to obtain keyword vectors and tag vectors.

3. The method for evolving a credit business system based on large-scale model (RAG) technology according to claim 2, characterized in that: The process of determining whether the response text can be retrieved from the prior knowledge base includes the following steps: Obtain the question vector; Match the question vector in the hot knowledge texts of the pre-existing database, and determine whether the answer text exists in the pre-existing database according to the preset similarity threshold. If it exists, return the answer text and obtain the topic tags of the topic knowledge base corresponding to the answer text. If it does not exist, the subject knowledge base will be used for retrieval.

4. The method for evolving a credit business system based on large-scale model (RAG) technology according to claim 3, characterized in that: The process of inputting the response text or standard text into the large model to obtain the result text includes: To input the response text into the large model and obtain the result text, the following steps are required: Obtain the question text, extract the topic keywords from the question text, match keywords based on the topic keywords, and obtain the corresponding topic tags; By combining topic tags, an output template for human input is obtained. The response text and prompt words are simultaneously input into the large model, and the result text is output through the output template. To input standard text into a large model and obtain the resulting text, the following steps are required: Obtain the question text, extract the topic keywords from the question text, match keywords based on the topic keywords, and obtain the corresponding topic tags; By combining topic tags, an output template for manual input is obtained, and knowledge text segments are acquired. These knowledge text segments and prompt words are then simultaneously input into the large model, and the result text is output through the output template.

5. The method for evolving a credit business system based on large-scale model (RAG) technology according to claim 4, characterized in that: Before determining whether a response text exists in the prior knowledge base, the following steps are required: Obtain feedback information and determine whether the feedback information is positive or negative; If the feedback is positive, then check the prior knowledge base to see if the question text exists; If the feedback is negative, then error correction information is obtained.

6. A credit business system evolution device based on large model RAG technology, characterized in that: include: The resource preprocessing module is used to acquire information resources, classify them, obtain multiple sets of topic resources, and simultaneously input these topic resources and manually input question-and-answer corpora into a large model. This outputs multiple topic texts, generates a topic knowledge base, refines the question-and-answer corpora, generates a pre-existing knowledge base, and associates the topic knowledge base with the pre-existing knowledge base. This includes: acquiring multiple sets of topic texts, vectorizing each set, generating the topic knowledge base; acquiring hot topic knowledge texts, constructing internal associations between the hot topic knowledge texts and the topic knowledge base; acquiring keywords and topic tags, associating keyword vectors and tag vectors with hot topic knowledge texts to generate the pre-existing knowledge base, and establishing a hot topic set, storing the hot topic knowledge texts in the hot topic set of the pre-existing knowledge base; establishing a topic set in the pre-existing knowledge base, storing keyword vectors and tag vectors in the topic set; and storing the topic knowledge base and the pre-existing knowledge base in a vector database. The input processing module is used to extract precise words from the question text and quantize the precise words to obtain the question vector. The data retrieval module determines whether the response text to the question vector can be retrieved from the pre-existing knowledge base. If so, the response text is obtained; otherwise, the standard text is obtained by retrieval through the pre-existing knowledge base and then through the topic knowledge base. This retrieval through the pre-existing knowledge base and then through the topic knowledge base includes: obtaining precise vocabulary; retrieving a topic set based on the precise vocabulary to obtain corresponding keyword vectors; obtaining topic tags based on the keyword vectors; obtaining the target topic knowledge base based on the topic tags; obtaining the question vector; retrieving the target topic knowledge base based on the question vector; obtaining knowledge text segments with similarity reaching a preset similarity threshold; and obtaining the topic text corresponding to the knowledge text segments to obtain the standard text. The results output module is used to input the response 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 feedback information input by humans based on the output result text, call the prior knowledge base, and retrieve the prior knowledge base; The feedback processing module determines whether a search result containing the response text exists in the prior knowledge base. If not, it updates the question vector, result text, and feedback information to the prior knowledge base. If the response text exists, it updates the prior knowledge base a second time, including: obtaining feedback information and determining whether it is positive or negative feedback; if it is positive feedback, obtaining the question text, question vector, and result text, and updating them to the prior knowledge base; if it is negative feedback, performing the following steps: deleting the result text from the prior knowledge base; obtaining correction information, vectorizing the correction information to obtain a correction vector; and updating the question text and correction vector to the prior knowledge base.

7. A computing device, comprising a memory and a processor, characterized in that, The memory stores executable code, and when the processor executes the executable code, it implements the method described in claims 1-5.

Images