Method and device for updating RAG knowledge base of financial customer service and storage medium

Abstract

This invention discloses a method, apparatus, and storage medium for updating a financial customer service RAG knowledge base, comprising: acquiring real-time financial information pushed by a target data source; determining, based on the real-time financial information and the financial customer service RAG knowledge base to be updated, content to be updated in the financial customer service RAG knowledge base to be updated, and determining whether there is an information conflict between the content to be updated and the real-time financial information; if so, generating conflict point content and update suggestions for the conflict point content based on the content to be updated and the real-time financial information; acquiring multi-source real-time related information within a preset time period associated with the conflict point content, and determining, based on the multi-source real-time related information, whether it is necessary to update the financial customer service RAG knowledge base to be updated; if so, updating the conflict point content in the financial customer service RAG knowledge base based on the update suggestions; otherwise, prohibiting the updating of the financial customer service RAG knowledge base.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and in particular to a method, apparatus and storage medium for updating a financial customer service RAG knowledge base. Background Technology

[0002] With the demonstrating power of Large Language Models (LLMs) in semantic understanding and multi-turn dialogue tasks in financial customer service, their practical value has attracted significant attention. However, the inherent "hallucination" problem of LLMs—generating grammatically fluent but factually incorrect fictitious content—poses a serious threat to the compliance and reliability crucial to financial services. Retrieval-Augmented Generation (RAG) technology has therefore become a core solution for the financial industry. When a user asks a question, the RAG system retrieves relevant text fragments in real time from the financial institution's private knowledge base (such as credit policies, regulatory documents, product brochures, and market data) as factual basis for its response. Therefore, to improve the response effectiveness of the RAG system, its knowledge base needs to be updated.

[0003] Currently, the RAG knowledge base is typically updated manually. However, this method is time-consuming and labor-intensive, leading to update delays; moreover, due to varying skill levels or negligence among staff, errors in knowledge updates can occur, thus affecting the subsequent search performance of the RAG-based system. Summary of the Invention

[0004] This invention provides a method, apparatus, and storage medium for updating a financial customer service RAG knowledge base, which mainly improves the updating efficiency and accuracy of the financial customer service RAG knowledge base.

[0005] According to a first aspect of the present invention, a method for updating a financial customer service RAG knowledge base is provided, comprising:

[0006] Obtain real-time financial information pushed by the target data source, and extract information summary information from the real-time financial information;

[0007] The information summary information is matched with the content in the financial customer service RAG knowledge base to be updated, and based on the similarity matching results, candidate content to be updated is determined in the financial customer service RAG knowledge base to be updated.

[0008] Determine the overall correlation between the candidate content to be updated and the real-time financial information. Based on the overall correlation, select the content to be updated from the candidate content to be updated. Determine whether there is an information conflict between the content to be updated and the real-time financial information. If so, generate content with conflict points and update suggestions for the content with conflict points based on the content to be updated and the real-time financial information.

[0009] Obtain multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether the financial customer service RAG knowledge base to be updated needs to be updated. If so, update the content of the conflict point in the financial customer service RAG knowledge base based on the update suggestion; otherwise, prohibit updating the financial customer service RAG knowledge base.

[0010] Optionally, determining whether to update the financial customer service RAG knowledge base based on the multi-source real-time related information includes:

[0011] Based on the timeliness and category information of the multi-source real-time related information, the degree of conflict between the content to be updated and the multi-source real-time related information is determined.

[0012] Based on the conflict correlation degree, select strongly correlated information from the multi-source real-time correlated information and determine the degree of support of the strongly correlated information for the update suggestion;

[0013] Based on the level of support, the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion are determined in the strongly related information messages. Based on the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion, it is determined whether the financial customer service RAG knowledge base to be updated needs to be updated.

[0014] Optionally, determining the degree of conflict between the content to be updated and the multi-source real-time related information based on the timeliness and category information of the multi-source real-time related information includes:

[0015] Determine the semantic correlation between the multi-source real-time related information and the conflict point content in the content to be updated, and based on the semantic correlation, determine the initial conflict correlation degree between the multi-source real-time related information and the conflict point content;

[0016] Determine the information release time of the multi-source real-time related information, and based on the current time and the information release time, determine the timeliness weight corresponding to the timeliness, and determine the category weight corresponding to the category information;

[0017] Based on the timeliness weight and the category weight, a conflict correlation correction coefficient is determined, and the initial conflict correlation is corrected based on the conflict correlation correction coefficient. The corrected initial conflict correlation is then used as the conflict correlation between the content to be updated and the multi-source real-time related information.

[0018] Optionally, determining the comprehensive correlation between the candidate content to be updated and the real-time financial information includes:

[0019] Determine the domain relevance, content relevance, and key information relevance between the candidate content to be updated and the real-time financial information;

[0020] Based on the domain relevance, the content relevance, and the key information relevance, a comprehensive relevance is determined;

[0021] Determining the domain relevance between the candidate content to be updated and the real-time financial information includes:

[0022] The domain keywords corresponding to the real-time financial information are determined, and the domain knowledge graph of the candidate content to be updated is determined, wherein the nodes in the domain knowledge graph are domain entities in the candidate content to be updated, and the edges between the nodes are the influence relationships between the domain entities;

[0023] The domain keywords are matched with each node in the domain knowledge graph for similarity. Based on the similarity matching results, target nodes similar to the domain keywords are determined in each node.

[0024] In the domain knowledge graph, the connection relationship between the target node and the remaining nodes is determined. Based on the connection relationship, the importance of the target node is determined. Based on the importance, the domain relevance between the candidate content to be updated and the real-time financial information is determined. The remaining nodes are each node in the domain knowledge graph after removing the target node.

[0025] Determining the content correlation between the candidate content to be updated and the real-time financial information includes:

[0026] The real-time financial information and the content to be updated are acquired in multimodal data, wherein the multimodal data includes at least two types of data, namely text data, time-series structured data, and chart data.

[0027] The information feature vectors of the multimodal data corresponding to the real-time financial information and the content feature vectors of the multimodal data corresponding to the content to be updated are determined respectively. The information feature vectors of each modality are cross-processed to obtain information cross feature vectors, and the content feature vectors of each modality are cross-processed to obtain content cross feature vectors.

[0028] Based on the information cross feature vector and the content cross feature vector, the content correlation degree between the candidate content to be updated and the real-time financial information is determined;

[0029] Determining the key information correlation between the candidate content to be updated and the real-time financial information includes:

[0030] Extract key content information from the candidate content to be updated and key information from the real-time financial information, and determine the information similarity between the key content information and the key information. Based on the information similarity, determine the key information correlation between the candidate content to be updated and the real-time financial information.

[0031] Optionally, determining the comprehensive relevance based on the domain relevance, the content relevance, and the key information relevance includes:

[0032] Obtain a sample dataset, wherein the sample dataset includes the sample neighborhood correlation degree, sample content correlation degree, and sample key information correlation degree between sample knowledge and sample information with knowledge update decision labels;

[0033] The sample neighborhood correlation degree, the sample content correlation degree, the sample key information correlation degree, and the initial weight coefficients corresponding to the sample neighborhood correlation degree, the sample content correlation degree, and the sample key information correlation degree are respectively input into the preset update decision prediction model to make the update decision prediction of the sample knowledge, so as to obtain the predicted update decision, and the prediction loss function of the preset update decision prediction model is determined based on the knowledge update decision label and the predicted update decision.

[0034] The initial weight coefficients corresponding to the sample neighborhood correlation, sample content correlation, and sample key information correlation are iteratively updated until the prediction loss function corresponding to the iteratively updated initial weight coefficients meets the loss requirements. Based on the initial weight coefficients after the last iteration, the domain correlation, content correlation, and key information correlation are weighted and summed to obtain the comprehensive correlation.

[0035] Optionally, selecting content to be updated from the candidate content based on the comprehensive relevance includes:

[0036] Select initial content to be updated from the candidate content to be updated that has a comprehensive relevance greater than a preset relevance threshold;

[0037] The initial content to be updated and the real-time financial information are input into a preset correlation prediction model to re-predict the correlation, thereby obtaining the rearranged correlation between the initial content to be updated and the real-time financial information. The preset correlation prediction model is trained in advance based on a dataset consisting of sample knowledge content and sample information with correlation labels.

[0038] Based on the rearranged relevance, a preset number of initial content to be updated are selected from the initial content to be updated as the content to be updated.

[0039] Optionally, before updating the content related to the conflict points in the financial customer service RAG knowledge base based on the update recommendations, the method further includes:

[0040] Multiple update recommendations were identified and evaluated for the terminal.

[0041] The update suggestion evaluation criteria information, which includes the real-time financial information and the content to be updated, is sent as reference information to each update suggestion evaluation terminal so that each update suggestion evaluation terminal can perform a zero-interaction evaluation of the update suggestion based on the reference information and obtain a zero-interaction evaluation result.

[0042] The zero-interaction evaluation result is received and sent to each of the update suggestion evaluation terminals respectively, so that each of the update suggestion evaluation terminals can perform terminal interaction evaluation on the update suggestion based on the zero-interaction evaluation result and the reference information to obtain the interaction evaluation result;

[0043] Receive the interaction evaluation result, and determine the verification result of the update suggestion based on the interaction evaluation result;

[0044] The update of the conflict point content in the financial customer service RAG knowledge base based on the update suggestion includes:

[0045] The content related to the conflict points in the financial customer service RAG knowledge base will be updated based on the verified update recommendations.

[0046] According to a second aspect of the present invention, a financial customer service RAG knowledge base updating device is provided, comprising:

[0047] The acquisition unit is used to acquire real-time financial information pushed by the target data source and extract information summary information from the real-time financial information.

[0048] The matching unit is used to perform similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and to determine candidate content to be updated in the financial customer service RAG knowledge base to be updated based on the similarity matching results.

[0049] The determining unit is used to determine the comprehensive correlation between the candidate content to be updated and the real-time financial information. Based on the comprehensive correlation, it selects the content to be updated from the candidate content to be updated and determines whether there is an information conflict between the content to be updated and the real-time financial information. If so, it generates conflict point content and update suggestions for the conflict point content based on the content to be updated and the real-time financial information.

[0050] The update unit is used to acquire multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether the financial customer service RAG knowledge base to be updated needs to be updated. If so, the content of the conflict point in the financial customer service RAG knowledge base is updated based on the update suggestion; otherwise, updating the financial customer service RAG knowledge base is prohibited.

[0051] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described financial customer service RAG knowledge base update method.

[0052] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described financial customer service RAG knowledge base update method.

[0053] According to the present invention, a method, apparatus, and storage medium for updating a financial customer service RAG knowledge base, compared with the current method of manually updating the RAG knowledge base, the present invention updates the knowledge base using real-time financial information, which can ensure the timeliness of the knowledge in the knowledge base; through similarity matching and comprehensive correlation analysis between real-time financial information and the RAG knowledge base, the present invention can accurately identify the knowledge in the RAG knowledge base that needs to be updated, thereby improving the accuracy of knowledge base updates; based on the comprehensive analysis of multi-source real-time related information associated with conflict points, the support level for updating the RAG knowledge base is determined, ensuring that the information in the knowledge base is always synchronized with market dynamics, thereby improving the accuracy of updating the knowledge in the RAG knowledge base; and by automating the knowledge base update, the update efficiency can be improved. Attached Figure Description

[0054] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0055] Figure 1 This invention provides a flowchart of a financial customer service RAG knowledge base update method according to an embodiment of the present invention.

[0056] Figure 2 This invention provides a flowchart of another financial customer service RAG knowledge base update method according to an embodiment of the present invention.

[0057] Figure 3 This diagram illustrates the structure of a financial customer service RAG knowledge base update device according to an embodiment of the present invention.

[0058] Figure 4 This invention provides a schematic diagram of another financial customer service RAG knowledge base updating device according to an embodiment of the present invention.

[0059] Figure 5 A schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0060] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0061] Currently, manually updating the RAG knowledge base is time-consuming and labor-intensive, leading to update delays; moreover, due to the varying technical skills of staff or negligence, errors in knowledge updates may occur.

[0062] To address the aforementioned problems, embodiments of the present invention provide a method for updating a financial customer service RAG knowledge base, such as... Figure 1 As shown, the method includes:

[0063] 101. Obtain real-time financial information pushed by the target data source and extract the information summary information from the real-time financial information.

[0064] The target data source can be at least one of the following: news, public accounts, self-media, etc.; real-time financial information includes hot news, policies, regulations, activities and announcements, such as interest rates, exchange rates, and regulatory effective dates in the financial field.

[0065] In this embodiment of the invention, the company utilizes its internal multi-source real-time information service to continuously push high-interest news, announcements, policies, and market activities. The real-time financial information is provided by the company's suppliers from multiple sources, and all content complies with the company's compliance requirements for customer presentation. After receiving the pushed real-time financial information, the server sends the information to a large financial model to generate a summary. This embodiment of the invention uses real-time financial information to update the knowledge base, ensuring the timeliness of the knowledge in the knowledge base.

[0066] 102. Perform similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and determine the candidate content to be updated in the financial customer service RAG knowledge base based on the similarity matching results.

[0067] In this embodiment of the invention, after obtaining the information summary information, the information summary information is vectorized to obtain information vectors, and the vector data is stored in a value database for subsequent similarity matching. Then, each knowledge content in the financial customer service RAG knowledge base to be updated is vectorized to obtain content vectors. Based on the information vectors and content vectors, the similarity between the information summary information and the content in the financial customer service RAG knowledge base to be updated is calculated using methods such as cosine similarity and Euclidean distance. Content with the highest similarity ranking in the RAG knowledge base is selected as candidate content to be updated. For example, the top 20 similar content in the RAG knowledge base is selected as candidate content to be updated. The similarity matching in this embodiment of the invention can accurately locate relevant paragraphs in the knowledge base, ensuring that the updated content highly matches user needs and improving the accuracy of knowledge base updates. In another embodiment of the present invention, the number of candidate content to be updated can be increased. For example, keywords in real-time financial information can be sent to a financial knowledge graph in a knowledge base. Based on graph structure analysis and semantic similarity evaluation, the top 20 most similar terms can be selected as candidate content to be updated. Fuzzy matching can be performed in the knowledge base using the keywords of real-time financial information. The correlation between knowledge and information can be calculated based on the keyword weights, and the top 20 content with the highest correlation can be obtained as candidate content to be updated.

[0068] 103. Determine the overall correlation between the candidate content to be updated and the real-time financial information. Based on the overall correlation, select the content to be updated from the candidate content to be updated and determine whether there is any information conflict between the content to be updated and the real-time financial information. If so, generate the content of the conflict and update suggestions for the content of the conflict based on the content to be updated and the real-time financial information.

[0069] In this embodiment of the invention, after determining the candidate content to be updated, in order to improve the accuracy of determining the content to be updated, it is also necessary to determine the domain relevance, content relevance, and key information relevance between the candidate content to be updated and the real-time financial information. Based on the domain relevance, content relevance, and key information relevance, a comprehensive relevance is determined. Finally, based on the comprehensive relevance, the final content to be updated is selected from the candidate content to be updated. This embodiment of the invention selects the final content to be updated by comprehensively analyzing the domain relevance, content relevance, and key information relevance, which can improve the accuracy of determining the content to be updated, thereby improving the accuracy of subsequent knowledge base updates. Further, the content to be updated and the real-time financial information are filled into the prompt word template and sent to the logical reasoning model. The reasoning model determines whether there is a logical or informational conflict between the content to be updated and the information expressed in the real-time financial information. If a conflict exists, an update suggestion is generated based on the real-time financial information (the information content is up-to-date; if a conflict exists, it means that the knowledge base content may contain outdated information) and the content to be updated, and the conflict point content and update suggestion are returned. For example, the content to be updated is: the 12-month installment fee rate for credit cards is 1.2% per installment. Real-time information indicates that the 12-month installment fee rate for credit cards has been reduced from 1.2% per installment to 0.9% per installment. The conflict between the content to be updated and the real-time financial information lies in the fee rate (1.2% vs 0.9%) and the effective date (May 1, 2024 vs. no specific date in the content to be updated). Based on these conflicts, the update suggestion is to update the 12-month installment fee rate for credit cards to 0.9% per installment. This invention, by automatically identifying conflict points and generating update suggestions, can improve the efficiency and accuracy of knowledge base updates.

[0070] 104. Obtain multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether it is necessary to update the financial customer service RAG knowledge base to be updated. If so, update the content of the conflict point in the financial customer service RAG knowledge base based on the update suggestion; otherwise, prohibit updating the financial customer service RAG knowledge base.

[0071] The preset time period is set according to actual needs, such as within 24 hours; multi-source real-time related information refers to real-time consultation information related to the content of the conflict point obtained from multiple data sources, including but not limited to news, announcements from authoritative institutions, public accounts, and self-media.

[0072] In this embodiment of the invention, multi-source real-time related information information associated with the content of the conflict point is obtained, and the degree of support of the multi-source real-time related information information for the conflict point update suggestion is statistically analyzed. For example, if the conflict point is the handling fee rate (real-time consultation information 1.2% vs. content to be updated 0.9%), and most of the handling fee rates in the multi-source real-time related information are 1.2%, then it is determined that the multi-source real-time related information information has a high degree of support for the conflict point update suggestion, and it is determined that the financial customer service RAG knowledge base to be updated needs to be updated. At this time, based on the update suggestion, the conflict point of the content to be updated is updated, such as adjusting 1.2% to 0.9%. Otherwise, it is determined that the financial customer service RAG knowledge base to be updated does not need to be updated. This embodiment of the invention verifies the update support of conflict content by introducing multi-source real-time related information information within a preset time period. Single information may contain typos or be delayed. Multi-source verification can significantly improve the accuracy of the updated content, avoid update deviations caused by erroneous information, and ensure that the information in the knowledge base is always synchronized with market dynamics.

[0073] Furthermore, after the update is complete, to further ensure accuracy, a review message for the knowledge base content update is pushed to all consultants (consultant side) and administrators (administrator side). During real-time assistance, visitor messages trigger real-time assistance Q&A logic. The assistance Q&A program searches the knowledge base, system pre-stored content, and personal pre-stored content based on the user's question and context. Knowledge that has been automatically updated and knowledge requiring manual review will display relevant judgment criteria, providing authorized consultants with secondary review and answer reference. If a consultant or administrator determines that the updated content of the large model is incorrect after reviewing the relevant information, they can immediately roll back the knowledge base or manually update the knowledge base; if they feel that the update suggestions submitted for review of the large model are acceptable, they can apply them with one click or apply them after modification.

[0074] According to the present invention, a method for updating a financial customer service RAG knowledge base, compared with the current method of manually updating the RAG knowledge base, uses real-time financial information to update the knowledge base, which can ensure the timeliness of the knowledge in the knowledge base; through similarity matching and comprehensive correlation analysis between real-time financial information and the RAG knowledge base, the knowledge in the RAG knowledge base that needs to be updated can be accurately identified, thereby improving the accuracy of knowledge base updates; based on the comprehensive analysis of multi-source real-time related information associated with conflict points, the degree of support for updating the RAG knowledge base is determined, ensuring that the information in the knowledge base is always synchronized with market dynamics, thereby improving the accuracy of updating the knowledge in the RAG knowledge base; and by automating the knowledge base update, the update efficiency can be improved. Therefore, this invention uses real-time financial information to update the knowledge base, ensuring the timeliness of the knowledge in the knowledge base. By matching the similarity between real-time financial information and the RAG knowledge base and conducting comprehensive correlation analysis, it can accurately identify the knowledge in the RAG knowledge base that needs to be updated, thereby improving the accuracy of knowledge base updates. Based on a comprehensive analysis of multi-source real-time related information associated with conflict points, it determines the degree of support for updating the RAG knowledge base, ensuring that the information in the knowledge base is always synchronized with market dynamics, thereby improving the accuracy of updating knowledge in the RAG knowledge base. By implementing knowledge base updates in an automated manner, it can improve the efficiency of knowledge base updates.

[0075] Furthermore, to better illustrate the process of updating the financial customer service RAG knowledge base described above, as a refinement and extension of the above embodiments, this invention provides another method for updating the financial customer service RAG knowledge base, such as... Figure 2 As shown, the method includes:

[0076] 201. Obtain real-time financial information pushed by the target data source and extract the information summary information from the real-time financial information.

[0077] Specifically, the target data sources cooperating with the company push financial information to the company in real time, and extract summary information from the financial information.

[0078] 202. Perform similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and determine the candidate content to be updated in the financial customer service RAG knowledge base based on the similarity matching results.

[0079] Specifically, the information summary information and the content in the financial customer service RAG knowledge base to be updated are vectorized respectively. Then, based on the vectorization results, the similarity between the information summary information and each piece of content in the RAG knowledge base is calculated. Finally, the content with a similarity greater than the preset similarity threshold (where the preset similarity threshold is set according to actual needs) is selected as the candidate content to be updated.

[0080] 203. Determine the overall correlation between the candidate content to be updated and the real-time financial information. Based on the overall correlation, select the content to be updated from the candidate content to be updated and determine whether there is any information conflict between the content to be updated and the real-time financial information. If so, generate the content of the conflict and update suggestions for the content of the conflict based on the content to be updated and the real-time financial information.

[0081] In this embodiment of the invention, to accurately determine the content to be updated, it is first necessary to determine the comprehensive correlation between the candidate content to be updated and the real-time financial information. Based on this, step 203 specifically includes: determining the domain correlation, content correlation, and key information correlation between the candidate content to be updated and the real-time financial information; and determining the comprehensive correlation based on the domain correlation, content correlation, and key information correlation. Specifically, determining the domain correlation between the candidate content to be updated and the real-time financial information includes: determining the domain keywords corresponding to the real-time financial information and determining the domain knowledge graph of the candidate content to be updated. The domain knowledge graph is used to determine the domain entities in the candidate content to be updated, where nodes represent domain entities and edges between nodes represent influence relationships between these entities. The domain keywords are matched with each node in the domain knowledge graph for similarity. Based on the similarity matching results, target nodes similar to the domain keywords are identified within each node. The connection relationships between the target nodes and the remaining nodes in the domain knowledge graph are determined. Based on these connection relationships, the importance of the target nodes is determined, and based on this importance, the domain relevance between the candidate content to be updated and the real-time financial information is determined. The remaining nodes... The nodes are the nodes in the domain knowledge graph after removing the target node; determining the content correlation between the candidate content to be updated and the real-time financial information includes: acquiring multimodal data of the real-time financial information and the content to be updated, wherein the multimodal data includes at least two types of data, such as text data, time-series structured data, and chart data; determining the information feature vector of the multimodal data corresponding to the real-time financial information and the content feature vector of the multimodal data corresponding to the content to be updated, and performing cross-processing on the information feature vector of each modality to obtain an information cross-feature vector, and determining the content feature vector of each modality. The feature vectors are cross-processed to obtain content cross-feature vectors; based on the information cross-feature vectors and the content cross-feature vectors, the content correlation between the candidate content to be updated and the real-time financial information is determined; the key information correlation between the candidate content to be updated and the real-time financial information is determined, including: extracting content key information from the candidate content to be updated and information key information from the real-time financial information, and determining the information similarity between the content key information and the information key information, and based on the information similarity, determining the key information correlation between the candidate content to be updated and the real-time financial information.

[0082] Specifically, real-time financial information text undergoes word segmentation, part-of-speech tagging, and stop word removal. For example, the information "The central bank announced a 5 basis point reduction in the LPR interest rate" is segmented into "Central Bank / announced / reduction / LPR interest rate / 5 basis points," retaining nouns, verbs, and financial terminology (such as "LPR interest rate" and "monetary policy"). Weighting: Keyword weights are calculated based on term frequency (TF) and inverse document frequency (IDF), and the words with the highest weights are selected as domain keywords. For example, in the above information, "LPR interest rate" (weight 0.8) and "reduction" (weight 0.6) are selected as domain keywords. Domain entities (such as financial products, institutions, and policies) are identified from candidate content to be updated. For example, in the candidate content to be updated, "A bank launched a wealth management product with an R3 risk level and an annualized return of 4%", the entities are "bank," "R3 risk level," "wealth management product," and "annualized return." Relationship labeling: Influence relationships between entities are defined (such as "belongs to," "influences," and "depends on"), and the direction of the edges is labeled. For example: "Wealth Management Products" → (belongs to) → "Bank"; "R3 Risk Level" → (impacts) → "Wealth Management Products"; "Annualized Rate of Return" → (depends on) → "Wealth Management Products". Graph Storage: Entities are used as nodes and relationships as edges to construct a directed graph structure, i.e., a domain knowledge graph. Domain keywords and knowledge graph node text are converted into vectors (e.g., using Word2Vec or BERT models). For example, the vector for the keyword "LPR interest rate" is [0.2, -0.5, 0.8], and the vector for the node "loan interest rate" is [0.1, -0.4, 0.7]. Based on these two vectors, cosine similarity is used to calculate the similarity between the domain keywords and each node in the domain knowledge graph. Nodes with similarity greater than a preset similarity threshold set according to actual needs are retained as target nodes. For example, if the keyword {LPR interest rate, reduction} matches the target node {loan interest rate, interest rate adjustment}. Next, the number of incoming edges (edges pointing to the node) and outgoing edges (edges pointing from the node) of the target node are calculated. For example, the node "loan interest rate" has an in-degree of 2 (from "central bank policy" and "market supply and demand") and an out-degree of 1 (pointing to "mortgage cost"). If the target node is located on multiple shortest paths, its importance is higher. For example, "loan interest rate" is a necessary node from "central bank policy" to "mortgage cost", so its centrality is high. Finally, the importance score of the target node is assigned by combining the in-degree, out-degree, and intermediary centrality. For example, the score of "loan interest rate" is 0.9. Finally, if the target node is unique, the domain relevance is equal to the importance score of the target node. If there are multiple target nodes (such as keywords matching {loan interest rate, interest rate adjustment}), the domain relevance is equal to the average of the importance scores of each node.

[0083] Furthermore, at least two modalities of data, such as text, time series, and charts, are extracted from real-time financial information and content to be updated. For example, structured and unstructured data are obtained from news websites, official announcements, and social media. For instance, a piece of real-time financial information might contain multimodal data such as a title text ("The central bank announced a 0.5% reduction in the reserve requirement ratio"), publication time (March 15, 2024, 10:00 AM), and accompanying charts (interest rate trend line charts). Content to be updated: For example, an entry in the knowledge base about the "reserve requirement ratio" might contain multimodal data such as a text description ("The reserve requirement ratio is the percentage of commercial banks' required reserves"), historical adjustment time series (adjustment records from Q1 to Q4 of 2023), and related charts (heatmaps showing the impact of reserve requirement ratio cuts on the stock market). Then, the data from different modalities are converted into computable numerical vectors. Subsequently, the information feature vectors of each modality are cross-processed, as are the content feature vectors of each modality. The specific cross-processing method is as follows: For each modality's information feature vector, perform feature-level cross-processing to obtain an information feature cross vector; perform element-level cross-processing to obtain an information element cross vector; perform low-order cross-processing to obtain a low-order cross vector; and transform the information feature cross vector, information element cross vector, and low-order cross-processing to obtain an information cross feature vector. Similarly, for each modality's content feature vector, the cross-processing method is as follows: For each modality's content feature vector, perform feature-level cross-processing to obtain a content feature cross vector; perform element-level cross-processing to obtain a content element cross vector; perform low-order cross-processing to obtain a low-order cross vector; and transform the content feature cross vector, content element cross vector, and low-order cross-processing to obtain a content cross feature vector.Specifically, taking the cross-processing of information feature vectors for each modality as an example, in order to fully utilize the relationships between data, extract more latent features, and simultaneously consider both high-order and low-order processing to make data utilization more efficient and the subsequent update results more accurate, meeting the needs of practical application scenarios, it is necessary to perform cross-processing on the information feature vectors of each modality. The specific cross-processing method is as follows: if the information feature vectors of each modality are (a1, a2), (b1, b2), and (c1, c2), the specific cross-processing methods include: performing feature-level cross-enhancement between different feature vectors, that is, after performing a Hadamard product on all elements of the vectors, performing a convolution transformation under a certain weight w1, obtaining the information feature cross vector f(w1×(a1×b1×c1, a2×b2×c2)); simultaneously, performing element-level cross-processing on all feature vector data, that is, after performing a Hadamard product on each element of the vectors, assigning different weight values ​​w2 and w3 to each product result, and then performing a linear transformation, obtaining the information element cross vector f(w2×a1×b1×c1, a2×b2×c2)). b1×c1, w3×a2×b2×c2); In addition, all feature vectors undergo low-order cross-enhancement processing, and the result of the cross-enhancement processing is assigned a weight coefficient w4, and then a linear transformation is performed to obtain the information low-order cross vector f(w4(a1,a2,b1,b2,c1,c2)); Finally, the above information feature cross vector, information element cross vector, and information low-order cross vector are transformed, such as by horizontal concatenation, to obtain the information cross feature vector. It should be noted that the above examples are only illustrative and do not limit the embodiments of this application. Thus, by performing cross-processing on feature vectors of different modalities, different features can be automatically or explicitly combined to generate new feature combinations. These combined features may contain complex nonlinear relationships between the original features, which can capture more refined and richer information in the data. That is, it can make full use of the relationships between various data, extract more latent features, and take into account both high-order and low-order processing, so that the data is utilized more fully and the subsequent knowledge base update results are more accurate, meeting the needs of practical application scenarios. After determining the information cross feature vector and the content cross feature vector, the similarity between the candidate content to be updated and the real-time financial information is calculated based on the information cross feature vector and the content cross feature vector, and this similarity is used as the content relevance.

[0084] Furthermore, core information is precisely extracted from candidate content to be updated and real-time financial information to provide a foundation for subsequent correlation calculations. If the content to be updated and the real-time consultation information are structured data (such as database entries or tables), key fields are directly extracted. If they are text descriptions (such as research reports or historical analyses), named entity recognition (NER) and keyword extraction techniques are used to extract key information. Then, the expression format of key information in both is standardized to eliminate semantic ambiguity. Similarity is calculated for each key information field (such as subject, operation, and time), and weights are assigned according to field importance to calculate the total similarity. Finally, this total similarity is used as the correlation score for key information.

[0085] Furthermore, after determining the domain relevance, content relevance, and key information relevance between candidate content to be updated and real-time financial information, it is also necessary to determine the comprehensive relevance based on the domain relevance, content relevance, and key information relevance. Based on this, the method includes: acquiring a sample dataset, wherein the sample dataset includes sample neighborhood relevance, sample content relevance, and sample key information relevance between sample knowledge and sample information with knowledge update decision labels; and assigning the sample neighborhood relevance, the sample content relevance, the sample key information relevance, and the initial values ​​corresponding to the sample neighborhood relevance, the sample content relevance, and the sample key information relevance, respectively. The weight coefficients are input into a preset update decision prediction model to predict the update decision of the sample knowledge, thereby obtaining the predicted update decision. Based on the knowledge update decision label and the predicted update decision, the prediction loss function of the preset update decision prediction model is determined. The initial weight coefficients corresponding to the sample neighborhood correlation, the sample content correlation, and the sample key information correlation are iteratively updated until the prediction loss function corresponding to the iteratively updated initial weight coefficients meets the loss requirement. Based on the initial weight coefficients after the last iteration update, the domain correlation, the content correlation, and the key information correlation are weighted and summed to obtain the comprehensive correlation.

[0086] Specifically, during model training, a pre-defined initial update decision prediction model is first constructed. Next, a training sample dataset is obtained. The dataset is ensured to contain all necessary files, and the data is converted to a format understandable by the pre-defined initial update decision prediction model. Finally, the model is trained and tested. Specifically, the dataset can be divided first: using randomness or a specific strategy (such as stratified sampling), the sample dataset is divided into training and test sets. The model is then trained using the training set, and tested using the test set to evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the model performance does not meet requirements, it can return to the training phase for further iterations or adjustments. This process yields a pre-defined update decision prediction model that meets the requirements. Furthermore, the model acquires the neighborhood correlation, content correlation, and key information correlation between sample knowledge and real-time information in a sample knowledge base with decision labels indicating whether to update or not. Using these correlations, with initial weight coefficients set according to actual needs, as input, and the update decision of sample knowledge as a prediction magic table, the model automatically learns the optimal weight coefficients for each correlation through training, aiming to minimize the loss function. Finally, the weights are normalized to ensure the sum of the weights is 1, resulting in the optimal combination of α, β, and γ values—the optimal combination of weight coefficients corresponding to neighborhood correlation, content correlation, and key information correlation, respectively. Finally, based on these weight coefficients, a weighted sum of the neighborhood correlation, content correlation, and key information correlation is calculated to obtain the comprehensive correlation between the candidate content to be updated and the real-time financial information.

[0087] Furthermore, after calculating the comprehensive correlation degree, it is necessary to select content to be updated from the candidate content to be updated based on the comprehensive correlation degree. Therefore, the method includes: selecting initial content to be updated from the candidate content to be updated whose comprehensive correlation degree is greater than a preset correlation threshold; inputting the initial content to be updated and the real-time financial information into a preset correlation degree prediction model to re-predict the correlation degree, obtaining a rearranged correlation degree between the initial content to be updated and the real-time financial information, wherein the preset correlation degree prediction model is pre-trained based on a dataset composed of sample knowledge content and sample information with correlation degree labels; and selecting a preset number of initial content to be updated as the content to be updated based on the rearranged correlation degree.

[0088] The preset relevance threshold is set according to actual needs. Specifically, to improve the prediction accuracy of the preset relevance prediction model, it is first necessary to train and build the preset relevance prediction model. Based on this, the method includes: during model training, firstly, building a preset initial relevance prediction model, and secondly, obtaining a sample dataset. Ensure the dataset contains all necessary files, including sample content to be updated with relevance labels and sample consultation information. Convert the data into a format that the preset initial relevance prediction model can understand, and finally train and test the model. Specifically, the dataset can be divided first: using random or specific strategies (such as stratified sampling), the sample dataset can be divided into a training set and a test set. Then, the model is trained using the training set, and the trained model is tested using the test set to evaluate its performance on unseen data. Precision, recall, and other metrics on the test set are calculated and recorded. If the model performance does not meet the requirements, it can return to the training phase for more iterations or adjustments. This yields a preset relevance prediction model that meets the requirements. The initial content to be updated and real-time financial information are then input into a preset correlation prediction model to re-predict the correlation, resulting in a rearranged correlation. A preset number of initial content items with a rearranged correlation greater than a preset rearrangement threshold are then selected as the content to be updated. The preset number and preset rearrangement threshold are set according to actual needs. This embodiment of the invention, through the prediction of rearranged correlation, can more accurately select the content to be updated, thereby improving the accuracy of subsequent knowledge base updates.

[0089] 204. Obtain multi-source real-time related information within a preset time period associated with the content of the conflict point.

[0090] Specifically, for example, searching for relevant information within 24 hours in a news service based on the content of the conflict.

[0091] 205. Based on the timeliness and category information of multi-source real-time related information, determine the degree of conflict between the content to be updated and the multi-source real-time related information.

[0092] Among them, timeliness refers to the release time of multi-source real-time related information, and category information refers to the information source categories such as policies and regulations, personal media, etc.

[0093] In this embodiment of the invention, in order to accurately update knowledge, it is also necessary to determine the degree of conflict between the content to be updated and the multi-source real-time related information. Based on this, step 205 includes: determining the semantic relationship between the multi-source real-time related information and the conflicting content in the content to be updated; determining the initial degree of conflict between the multi-source real-time related information and the conflicting content based on the semantic relationship; determining the information release time of the multi-source real-time related information; determining the timeliness weight corresponding to the timeliness based on the current time and the information release time; determining the category weight corresponding to the category information; determining the conflict correlation correction coefficient based on the timeliness weight and the category weight; correcting the initial degree of conflict based on the conflict correlation correction coefficient; and using the corrected initial degree of conflict as the degree of conflict between the content to be updated and the multi-source real-time related information.

[0094] Specifically, semantic relevance prediction models are used to predict the semantic relevance between multi-source real-time related information and conflict points in the content to be updated. The stronger the semantic relevance, the greater the initial conflict relevance between the multi-source real-time related information and the conflict points. The closer the information's publication time is to the current time, the greater the timeliness weight; for example, the timeliness weight coefficient... It can be calculated using the following formula:

[0095]

[0096] in, Current time For information release time, This is the maximum time difference set according to actual needs. The more authoritative the publishing platform of the category information, the greater its corresponding category weight. Then, the timeliness weight and the category weight are added together to obtain the conflict correlation correction coefficient. The conflict correlation correction coefficient is then multiplied by the initial conflict correlation to obtain the conflict correlation degree.

[0097] 206. Based on the degree of conflict correlation, select strongly correlated information from multi-source real-time correlated information and determine the degree of support of strongly correlated information for update suggestions.

[0098] In this embodiment of the invention, information with a conflict correlation greater than a preset conflict threshold (set according to actual needs) is selected from multi-source real-time related information as strongly related information. If the information in the strongly related information is the same as the information in the real-time financial information regarding the conflict point, i.e., if the conflict point is: transaction fee rate (real-time financial information 1.2% vs. content to be updated 0.9%), and the update suggestion is to adjust the transaction fee rate in the content to be updated from 0.9% to 1.2%, and the transaction fee rate of the strongly related information regarding the same content is also 1.2%, then its support for the update suggestion is determined to be 100%, and the strongly related information is judged to support the update suggestion. If the transaction fee rate of the strongly related information regarding the same content is also 0.8%, the strongly related information is judged not to support the update suggestion. Specifically, the strongly related information and the update suggestion can be input into a large model, which can output whether the strongly related information supports or does not support the update suggestion.

[0099] 207. Based on the level of support, determine the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion among the strongly related information messages. Based on the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion, determine whether the financial customer service RAG knowledge base needs to be updated. If so, update the conflict point content in the financial customer service RAG knowledge base according to the update suggestion. Otherwise, prohibit updating the financial customer service RAG knowledge base.

[0100] Specifically, the number of strongly related information pieces that support update suggestions and the number of information pieces that do not support update suggestions and do not support strongly related information pieces are determined. If the number of strongly related information pieces is greater than or equal to a first preset threshold, and the number of information pieces that do not support strongly related information pieces is less than or equal to a second preset threshold, then it is determined that the financial customer service RAG knowledge base to be updated needs to be updated; otherwise, it is determined that the financial customer service RAG knowledge base to be updated does not need to be updated. The first preset threshold is greater than the second preset threshold. For example, if the first preset threshold is 3 and the second preset threshold is 1, then when the number of strongly related information pieces... 3. And it does not support a large number of strongly related information items. If the number of strongly related information items is less than 3, or if the number of strongly related information items is greater than 1, then it is determined that the financial customer service RAG knowledge base to be updated needs to be updated.

[0101] Furthermore, if the knowledge base needs to be updated, the update suggestions also need to be evaluated to increase the accuracy of the update. Based on this, the method includes: determining multiple update suggestion evaluation terminals; sending update suggestion evaluation criteria information containing the real-time financial information and the content to be updated as reference information to each update suggestion evaluation terminal, so that each update suggestion evaluation terminal performs a zero-interaction evaluation of the update suggestion based on the reference information, and obtains a zero-interaction evaluation result; receiving the zero-interaction evaluation result and sending the zero-interaction evaluation result to each update suggestion evaluation terminal, so that each update suggestion evaluation terminal performs a terminal interaction evaluation of the update suggestion based on the zero-interaction evaluation result and the reference information, and obtains an interaction evaluation result; receiving the interaction evaluation result and determining the verification result of the update suggestion based on the interaction evaluation result; updating the conflict point content in the financial customer service RAG knowledge base based on the update suggestion includes: updating the conflict point content in the financial customer service RAG knowledge base based on the verified update suggestion.

[0102] Specifically, the various update suggestion evaluation terminals can be representative expert terminals from different fields; the update suggestion evaluation criteria information can include questionnaire information such as the items to be evaluated, the range of evaluation scores, weighting values, and evaluation precautions. For example, to facilitate the evaluation terminals in assigning (or scoring) update suggestions, a special scoring selection table needs to be prepared, along with detailed selection instructions, clarifying the standards and requirements for assignment. In the first round, the update suggestion evaluation criteria information and other reference information are sent to each update suggestion evaluation terminal. Each update suggestion evaluation terminal scores the update suggestions without communication based on the update suggestion evaluation criteria information. Then, the zero-interaction scoring results from each evaluation terminal are collected, and corresponding data is processed (such as calculating the average value, coefficient of variation, etc.). The processed results are then sent back to each evaluation terminal, allowing each evaluation terminal to perform interactive scoring assignments based on the processed results, thus obtaining interactive scoring results. Finally, the score corresponding to the update suggestion is determined based on the interactive scoring results. It should be noted that the update suggestion scoring process is not limited to the above two rounds of scoring; multiple rounds of scoring processes, such as three or five rounds, can also be performed. Finally, based on the final score of the update suggestion, the verification result of the update suggestion is determined. If the score is greater than the preset score threshold (the preset score threshold is set according to actual needs), the verification is deemed successful; otherwise, the verification fails. The content of conflict points in the knowledge base is updated based on the verified update suggestion. If the verification fails, a new update suggestion needs to be determined, and the conflict points are updated based on the newly determined update suggestion. This embodiment of the invention verifies update suggestions by combining zero-interaction scoring and interactive scoring. Since the first round of independent scoring requires the evaluation terminal to complete the assignment without communication, it effectively isolates group thinking and authority interference. The interactive stage allows the evaluation terminal to re-examine its own score based on the organized results (such as average values ​​and points of disagreement), correcting erroneous assignments caused by information misunderstanding or negligence, ultimately improving the match between the results and the true value.

[0103] According to another financial customer service RAG knowledge base update method provided by the present invention, compared with the current method of manually updating the RAG knowledge base, the present invention updates the knowledge base through real-time financial information, which can ensure the timeliness of the knowledge in the knowledge base; through similarity matching and comprehensive correlation analysis between real-time financial information and the RAG knowledge base, the knowledge that needs to be updated in the RAG knowledge base can be accurately identified, thereby improving the accuracy of knowledge base updates; based on the comprehensive analysis of multi-source real-time related information associated with conflict points, the degree of support for updating the RAG knowledge base is determined, ensuring that the information in the knowledge base is always synchronized with market dynamics, thereby improving the accuracy of updating the knowledge in the RAG knowledge base; and by automating the knowledge base update, the update efficiency of the knowledge base can be improved.

[0104] Furthermore, as Figure 1 In specific implementation, embodiments of the present invention provide a financial customer service RAG knowledge base update device, such as... Figure 3 As shown, the device includes: an acquisition unit 31, a matching unit 32, a determination unit 33, and an update unit 34.

[0105] The acquisition unit 31 can be used to acquire real-time financial information pushed by the target data source and extract information summary information from the real-time financial information.

[0106] The matching unit 32 can be used to perform similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and based on the similarity matching result, determine the candidate content to be updated in the financial customer service RAG knowledge base to be updated.

[0107] The determining unit 33 can be used to determine the comprehensive correlation between the candidate content to be updated and the real-time financial information. Based on the comprehensive correlation, it selects the content to be updated from the candidate content to be updated and determines whether there is an information conflict between the content to be updated and the real-time financial information. If so, it generates conflict point content and update suggestions for the conflict point content based on the content to be updated and the real-time financial information.

[0108] The update unit 34 can be used to obtain multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether the financial customer service RAG knowledge base to be updated needs to be updated. If so, the content of the conflict point in the financial customer service RAG knowledge base is updated based on the update suggestion; otherwise, updating the financial customer service RAG knowledge base is prohibited.

[0109] In specific application scenarios, to determine whether the RAG knowledge base for financial customer service needs to be updated, such as... Figure 4 As shown, the update unit 34 includes a determination module 341 and a judgment module 342.

[0110] The determining module 341 can be used to determine the degree of conflict between the content to be updated and the multi-source real-time related information based on the timeliness and category information of the multi-source real-time related information.

[0111] The determining module 341 can also be used to select strongly correlated information information from the multi-source real-time correlated information information based on the conflict correlation degree, and determine the degree of support of the strongly correlated information information for the update suggestion.

[0112] The judgment module 342 can be used to determine, based on the degree of support, the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion in the strongly related information messages, and to determine, based on the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion, whether the financial customer service RAG knowledge base to be updated needs to be updated.

[0113] In specific application scenarios, in order to determine the degree of conflict between the content to be updated and the multi-source real-time related information, the determining module 341 can be specifically used to determine the semantic relationship between the multi-source real-time related information and the conflicting content in the content to be updated; based on the semantic relationship, determine the initial degree of conflict between the multi-source real-time related information and the conflicting content; determine the information release time of the multi-source real-time related information; based on the current time and the information release time, determine the timeliness weight corresponding to the timeliness, and determine the category weight corresponding to the category information; based on the timeliness weight and the category weight, determine the conflict relationship correction coefficient, and correct the initial degree of conflict based on the conflict relationship correction coefficient, and use the corrected initial degree of conflict as the degree of conflict between the content to be updated and the multi-source real-time related information.

[0114] In specific application scenarios, to determine the comprehensive correlation between candidate content to be updated and real-time financial information, the determining unit 33 can specifically be used to determine the domain correlation, content correlation, and key information correlation between the candidate content to be updated and the real-time financial information; based on the domain correlation, content correlation, and key information correlation, a comprehensive correlation is determined; wherein, determining the domain correlation between the candidate content to be updated and the real-time financial information includes: determining the domain keywords corresponding to the real-time financial information, and determining the domain knowledge graph of the candidate content to be updated, wherein the domain... In the knowledge graph, nodes represent domain entities within the candidate content to be updated, and edges between nodes represent influence relationships between these domain entities. The domain keywords are matched for similarity with each node in the domain knowledge graph. Based on the similarity matching results, target nodes similar to the domain keywords are identified within each node. The connection relationships between the target nodes and the remaining nodes in the domain knowledge graph are determined. Based on these connection relationships, the importance of the target nodes is determined, and based on this importance, the domain relevance between the candidate content to be updated and the real-time financial information is determined. The remaining nodes are those after removing the relevant nodes. Each node in the domain knowledge graph following the target node; determining the content correlation between the candidate content to be updated and the real-time financial information, including: acquiring multimodal data of the real-time financial information and the content to be updated, wherein the multimodal data includes at least two types of data selected from text data, time-series structured data, and chart data; determining the information feature vector of the multimodal data corresponding to the real-time financial information and the content feature vector of the multimodal data corresponding to the content to be updated, and performing cross-processing on the information feature vector of each modality to obtain an information cross-feature vector, and processing the content feature vector of each modality... Cross-processing is performed to obtain content cross-feature vectors; based on the information cross-feature vectors and the content cross-feature vectors, the content correlation degree between the candidate content to be updated and the real-time financial information is determined; the key information correlation degree between the candidate content to be updated and the real-time financial information is determined, including: extracting content key information from the candidate content to be updated and information key information from the real-time financial information, and determining the information similarity between the content key information and the information key information, and based on the information similarity, determining the key information correlation degree between the candidate content to be updated and the real-time financial information.

[0115] In specific application scenarios, in order to determine the overall correlation degree, the determining unit 33 includes an acquisition module 331, a prediction module 332, and an update module 333.

[0116] The acquisition module 331 can be used to acquire a sample dataset, wherein the sample dataset includes the sample neighborhood correlation degree, sample content correlation degree, and sample key information correlation degree between sample knowledge and sample information with knowledge update decision labels.

[0117] The prediction module 332 can be used to input the sample neighborhood correlation degree, the sample content correlation degree, the sample key information correlation degree, and the initial weight coefficients corresponding to the sample neighborhood correlation degree, the sample content correlation degree, and the sample key information correlation degree into a preset update decision prediction model to make update decision prediction for the sample knowledge, obtain the predicted update decision, and determine the prediction loss function of the preset update decision prediction model based on the knowledge update decision label and the predicted update decision.

[0118] The update module 333 can be used to iteratively update the initial weight coefficients corresponding to the sample neighborhood correlation, the sample content correlation, and the sample key information correlation, until the prediction loss function corresponding to the iteratively updated initial weight coefficients meets the loss requirements. Based on the initial weight coefficients after the last iterative update, the domain correlation, the content correlation, and the key information correlation are weighted and summed to obtain the comprehensive correlation.

[0119] In specific application scenarios, in order to select the content to be updated from the candidate content to be updated, the determining unit 33 further includes a selection module 334.

[0120] The selection module 334 can be used to select initial content to be updated from the candidate content to be updated that has a comprehensive relevance greater than a preset relevance threshold.

[0121] The prediction module 332 can also be used to input the initial content to be updated and the real-time financial information into a preset correlation prediction model to re-predict the correlation, thereby obtaining the rearranged correlation between the initial content to be updated and the real-time financial information. The preset correlation prediction model is trained in advance on a dataset composed of sample knowledge content and sample information with correlation labels.

[0122] The selection module 334 can be specifically used to select a preset number of initial content to be updated as the content to be updated from the initial content to be updated, based on the rearranged correlation degree.

[0123] In specific application scenarios, the device further includes an evaluation unit 35 for evaluating the update suggestions.

[0124] The evaluation unit 35 can be used to determine multiple update suggestion evaluation terminals; send update suggestion evaluation criteria information containing the real-time financial information and the content to be updated as reference information to each update suggestion evaluation terminal, so that each update suggestion evaluation terminal can perform a zero-interaction evaluation of the update suggestion based on the reference information to obtain a zero-interaction evaluation result; receive the zero-interaction evaluation result and send the zero-interaction evaluation result to each update suggestion evaluation terminal, so that each update suggestion evaluation terminal can perform a terminal interaction evaluation of the update suggestion based on the zero-interaction evaluation result and the reference information to obtain an interaction evaluation result; receive the interaction evaluation result and determine the verification result of the update suggestion based on the interaction evaluation result.

[0125] In specific application scenarios, in order to update the knowledge base, the update unit 34 can be used to update the content of the conflict points in the financial customer service RAG knowledge base based on the verified update suggestions.

[0126] It should be noted that other corresponding descriptions of the functional modules involved in the financial customer service RAG knowledge base update device provided in this embodiment of the invention can be found in the following references. Figure 1 The corresponding description of the method shown will not be repeated here.

[0127] Based on the above, Figure 1 Correspondingly, this embodiment of the invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the following steps: acquiring real-time financial information pushed by a target data source and extracting information summary information from the real-time financial information; performing similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and determining candidate content to be updated in the financial customer service RAG knowledge base based on the similarity matching result; determining the comprehensive correlation between the candidate content to be updated and the real-time financial information, and, based on the comprehensive correlation, selecting the candidate content to be updated... The system selects content to be updated and determines whether there is an information conflict between the content to be updated and the real-time financial information. If so, it generates conflict point content and update suggestions for the conflict point content based on the content to be updated and the real-time financial information. It obtains multi-source real-time related information within a preset time period associated with the conflict point content and determines whether the financial customer service RAG knowledge base to be updated needs to be updated based on the multi-source real-time related information. If so, it updates the conflict point content in the financial customer service RAG knowledge base based on the update suggestions; otherwise, it prohibits updating the financial customer service RAG knowledge base.

[0128] Based on the above, Figure 1 The method shown and as Figure 3 The embodiment of the device shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 5 As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored on the memory 42 and executable on the processor. Both the memory 42 and the processor 41 are mounted on a bus 43. When the processor 41 executes the program, it performs the following steps: acquiring real-time financial information pushed by a target data source and extracting information summary information from the real-time financial information; performing similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and determining candidate content to be updated in the financial customer service RAG knowledge base based on the similarity matching results; determining the comprehensive correlation between the candidate content to be updated and the real-time financial information, and based on the... The system comprehensively assesses relevance, selects content to be updated from the candidate content to be updated, and determines whether there is an information conflict between the content to be updated and the real-time financial information. If so, it generates content related to the conflict point and update suggestions for the content related to the conflict point based on the content to be updated and the real-time financial information. It obtains multi-source real-time related information within a preset time period associated with the content related to the conflict point, and determines whether it is necessary to update the financial customer service RAG knowledge base to be updated based on the multi-source real-time related information. If so, it updates the content related to the conflict point in the financial customer service RAG knowledge base based on the update suggestions; otherwise, it prohibits updating the financial customer service RAG knowledge base.

[0129] Through the technical solution of this invention, the knowledge base is updated using real-time financial information, ensuring the timeliness of the knowledge in the knowledge base; by matching the similarity between real-time financial information and the RAG knowledge base and conducting comprehensive correlation analysis, the knowledge in the RAG knowledge base that needs to be updated can be accurately identified, thereby improving the accuracy of knowledge base updates; based on the comprehensive analysis of multi-source real-time related information associated with conflict points, the degree of support for updating the RAG knowledge base is determined, ensuring that the information in the knowledge base is always synchronized with market dynamics, thereby improving the accuracy of updating knowledge in the RAG knowledge base; and by implementing knowledge base updates in an automated manner, the update efficiency of the knowledge base can be improved.

[0130] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for updating a financial customer service RAG knowledge base, characterized in that, include: Obtain real-time financial information pushed by the target data source, and extract information summary information from the real-time financial information; The information summary information is matched with the content in the financial customer service RAG knowledge base to be updated, and based on the similarity matching results, candidate content to be updated is determined in the financial customer service RAG knowledge base to be updated. Determine the overall correlation between the candidate content to be updated and the real-time financial information. Based on the overall correlation, select the content to be updated from the candidate content to be updated. Determine whether there is an information conflict between the content to be updated and the real-time financial information. If so, generate content with conflict points and update suggestions for the content with conflict points based on the content to be updated and the real-time financial information. Obtain multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether the financial customer service RAG knowledge base to be updated needs to be updated. If so, update the content of the conflict point in the financial customer service RAG knowledge base based on the update suggestion; otherwise, prohibit updating the financial customer service RAG knowledge base.

2. The method according to claim 1, characterized in that, The step of determining whether the financial customer service RAG knowledge base to be updated needs to be updated based on the multi-source real-time related information includes: Based on the timeliness and category information of the multi-source real-time related information, the degree of conflict between the content to be updated and the multi-source real-time related information is determined. Based on the conflict correlation degree, select strongly correlated information from the multi-source real-time correlated information and determine the degree of support of the strongly correlated information for the update suggestion; Based on the level of support, the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion are determined in the strongly related information messages. Based on the number of strongly related information messages that support the update suggestion and the number of strongly related information messages that do not support the update suggestion, it is determined whether the financial customer service RAG knowledge base to be updated needs to be updated.

3. The method according to claim 2, characterized in that, The determination of the conflict correlation degree between the content to be updated and the multi-source real-time related information based on the timeliness and category information of the multi-source real-time related information includes: Determine the semantic correlation between the multi-source real-time related information and the conflict point content in the content to be updated, and based on the semantic correlation, determine the initial conflict correlation degree between the multi-source real-time related information and the conflict point content; Determine the information release time of the multi-source real-time related information, and based on the current time and the information release time, determine the timeliness weight corresponding to the timeliness, and determine the category weight corresponding to the category information; Based on the timeliness weight and the category weight, a conflict correlation correction coefficient is determined, and the initial conflict correlation is corrected based on the conflict correlation correction coefficient. The corrected initial conflict correlation is then used as the conflict correlation between the content to be updated and the multi-source real-time related information.

4. The method according to claim 1, characterized in that, Determining the comprehensive correlation between the candidate content to be updated and the real-time financial information includes: Determine the domain relevance, content relevance, and key information relevance between the candidate content to be updated and the real-time financial information; Based on the domain relevance, the content relevance, and the key information relevance, a comprehensive relevance is determined; Determining the domain relevance between the candidate content to be updated and the real-time financial information includes: The domain keywords corresponding to the real-time financial information are determined, and the domain knowledge graph of the candidate content to be updated is determined, wherein the nodes in the domain knowledge graph are domain entities in the candidate content to be updated, and the edges between the nodes are the influence relationships between the domain entities; The domain keywords are matched with each node in the domain knowledge graph for similarity. Based on the similarity matching results, target nodes similar to the domain keywords are determined in each node. In the domain knowledge graph, the connection relationship between the target node and the remaining nodes is determined. Based on the connection relationship, the importance of the target node is determined. Based on the importance, the domain relevance between the candidate content to be updated and the real-time financial information is determined. The remaining nodes are each node in the domain knowledge graph after removing the target node. Determining the content correlation between the candidate content to be updated and the real-time financial information includes: The real-time financial information and the content to be updated are acquired in multimodal data, wherein the multimodal data includes at least two types of data, namely text data, time-series structured data, and chart data. The information feature vectors of the multimodal data corresponding to the real-time financial information and the content feature vectors of the multimodal data corresponding to the content to be updated are determined respectively. The information feature vectors of each modality are cross-processed to obtain information cross feature vectors, and the content feature vectors of each modality are cross-processed to obtain content cross feature vectors. Based on the information cross feature vector and the content cross feature vector, the content correlation degree between the candidate content to be updated and the real-time financial information is determined; Determining the key information correlation between the candidate content to be updated and the real-time financial information includes: Extract key content information from the candidate content to be updated and key information from the real-time financial information, and determine the information similarity between the key content information and the key information. Based on the information similarity, determine the key information correlation between the candidate content to be updated and the real-time financial information.

5. The method according to claim 4, characterized in that, The determination of comprehensive relevance based on the domain relevance, the content relevance, and the key information relevance includes: Obtain a sample dataset, wherein the sample dataset includes the sample neighborhood correlation degree, sample content correlation degree, and sample key information correlation degree between sample knowledge and sample information with knowledge update decision labels; The sample neighborhood correlation degree, the sample content correlation degree, the sample key information correlation degree, and the initial weight coefficients corresponding to the sample neighborhood correlation degree, the sample content correlation degree, and the sample key information correlation degree are respectively input into the preset update decision prediction model to make the update decision prediction of the sample knowledge, so as to obtain the predicted update decision, and the prediction loss function of the preset update decision prediction model is determined based on the knowledge update decision label and the predicted update decision. The initial weight coefficients corresponding to the sample neighborhood correlation, sample content correlation, and sample key information correlation are iteratively updated until the prediction loss function corresponding to the iteratively updated initial weight coefficients meets the loss requirements. Based on the initial weight coefficients after the last iteration, the domain correlation, content correlation, and key information correlation are weighted and summed to obtain the comprehensive correlation.

6. The method according to claim 1, characterized in that, The step of selecting content to be updated from the candidate content based on the comprehensive relevance includes: Select initial content to be updated from the candidate content to be updated that has a comprehensive correlation degree greater than a preset correlation threshold; The initial content to be updated and the real-time financial information are input into a preset correlation prediction model to re-predict the correlation, thereby obtaining the rearranged correlation between the initial content to be updated and the real-time financial information. The preset correlation prediction model is trained in advance based on a dataset consisting of sample knowledge content and sample information with correlation labels. Based on the rearranged relevance, a preset number of initial content to be updated are selected from the initial content to be updated as the content to be updated.

7. The method according to claim 1, characterized in that, Before updating the content related to the conflict points in the financial customer service RAG knowledge base based on the update recommendations, the method further includes: Multiple update recommendations were identified and evaluated for the terminal. The update suggestion evaluation criteria information, which includes the real-time financial information and the content to be updated, is sent as reference information to each update suggestion evaluation terminal so that each update suggestion evaluation terminal can perform a zero-interaction evaluation of the update suggestion based on the reference information and obtain a zero-interaction evaluation result. The zero-interaction evaluation result is received and sent to each of the update suggestion evaluation terminals respectively, so that each of the update suggestion evaluation terminals can perform terminal interaction evaluation on the update suggestion based on the zero-interaction evaluation result and the reference information to obtain the interaction evaluation result; Receive the interaction evaluation result, and determine the verification result of the update suggestion based on the interaction evaluation result; The update of the conflict point content in the financial customer service RAG knowledge base based on the update suggestion includes: The content related to the conflict points in the financial customer service RAG knowledge base will be updated based on the verified update recommendations.

8. A financial customer service RAG knowledge base update device, characterized in that, include: The acquisition unit is used to acquire real-time financial information pushed by the target data source and extract information summary information from the real-time financial information. The matching unit is used to perform similarity matching between the information summary information and the content in the financial customer service RAG knowledge base to be updated, and to determine candidate content to be updated in the financial customer service RAG knowledge base to be updated based on the similarity matching results. The determining unit is used to determine the comprehensive correlation between the candidate content to be updated and the real-time financial information. Based on the comprehensive correlation, it selects the content to be updated from the candidate content to be updated and determines whether there is an information conflict between the content to be updated and the real-time financial information. If so, it generates conflict point content and update suggestions for the conflict point content based on the content to be updated and the real-time financial information. The update unit is used to acquire multi-source real-time related information within a preset time period associated with the content of the conflict point, and based on the multi-source real-time related information, determine whether the financial customer service RAG knowledge base to be updated needs to be updated. If so, the content of the conflict point in the financial customer service RAG knowledge base is updated based on the update suggestion; otherwise, updating the financial customer service RAG knowledge base is prohibited.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images