A cooperative bank customer manager scratch note quality detection system

Abstract

This invention discloses a collaborative quality inspection system for bank customer manager's notes. Deployed on the Dify platform, the system stores historical notes through a knowledge base management module and uses a semantic similarity detection engine to identify cross-document plagiarism. It employs a large language model for intelligent multi-customer identification and customer consistency verification, resolving issues such as multiple customers per note and irrelevance to the subject line. A weighted scoring card is used to integrate three dimensions of risk, enabling automated quality inspection decisions. This invention leverages the workflow orchestration capabilities of the Dify platform, integrating vector knowledge base semantic retrieval with large-scale model multi-task reasoning to achieve real-time, multi-dimensional, and automated quality control upon note submission, significantly improving the quality of bank customer insight data.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and artificial intelligence, specifically to a collaborative bank customer manager's note-taking quality inspection system. Background Technology

[0002] In daily customer marketing, post-loan management, and due diligence, bank account managers need to quickly record fragmented texts (commonly known as "notes") on mobile devices, including customer information, key communication points, and business progress. These records are crucial for bank customer management, risk control, and compliance audits. The quality of these notes directly impacts the accuracy of customer insights and the effectiveness of business decisions. However, current note management faces the following prominent problems: (1) Plagiarism is rampant: Account managers plagiarize each other's notes, resulting in serious homogenization of knowledge base data and loss of customer insight value. Traditional text plagiarism detection technology fails because it cannot identify semantic similarity. (2) Multiple customer confusion: Multiple customer visit details are incorrectly recorded in a single note, violating the "one customer, one note" management standard. Manual review is inefficient and prone to omissions. (3) The content is not related to the title: the content of the notes is seriously inconsistent with the customer indicated in the title. For example, the title is "Visiting Customer A", but the content records customer B or fabricated information. Traditional rule matching is difficult to identify deep semantic inconsistencies.

[0003] Existing technologies have the following shortcomings: (1) "Synonymous rewriting" copy-paste cannot be identified by rules or keyword matching; (2) Semantic errors such as multiple customers in a single record or customers not matching the text are difficult to detect; (3) The accuracy of traditional text similarity (edit distance, Jaccard, BM25) drops sharply under long texts and colloquial descriptions; (4) The bank’s internal product terminology and compliance expressions are updated frequently, and the detection model cannot be synchronized at low cost.

[0004] Simple text comparison for plagiarism prevention and consistency checks either rely on manual sampling and lack a systematic intelligent quality inspection solution. While large language models possess powerful semantic understanding capabilities, a single model cannot effectively solve the coordination problem between knowledge accumulation and real-time verification. Dify, as an open-source large model application development platform, provides knowledge base management and LLM orchestration capabilities, but there is still no complete quality inspection solution based on this platform specifically for the banking note-taking scenario. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a collaborative bank customer manager's note-taking quality inspection system.

[0006] To achieve the above objectives, the present invention provides a collaborative bank customer manager's note-taking quality inspection system, comprising: The real-time acquisition and preprocessing module provides a standardized input interface to receive the notes data uploaded by the account manager's terminal and preprocess the received notes data. The knowledge base management module, based on the Dify knowledge base, calls the Embedding model to vectorize and encode the preprocessed notes data, writes it into the vector database to build an index, and builds a hierarchical knowledge base according to business lines to realize the automatic entry of historical notes data into the database, and supports similarity retrieval and access permission isolation. The semantic similarity detection engine retrieves Top-K candidate results from the knowledge base through vector similarity, and then performs fine ranking by the Rerank model to comprehensively evaluate semantic relevance; by setting a similarity threshold λ, it identifies semantic plagiarism behavior across account managers and across time dimensions, and outputs semantic plagiarism detection results; The multi-customer intelligent identification module is used to define customer entity extraction functions, parse the text content of the preprocessed notes data, and output a list of customer names and confidence score detection results. The customer consistency verification module is used to determine whether the core action subject of the preprocessed notes data is consistent with the title "Customer", and outputs the customer consistency detection result. The quality inspection report generation module is used to aggregate the semantic plagiarism detection results, customer consistency detection results, and customer consistency detection results, and generate a structured quality inspection report.

[0007] Furthermore, the threshold λ is automatically adjusted based on the account manager's rank and business line.

[0008] Furthermore, the multi-customer intelligent identification module defines a customer entity extraction function based on the Function Calling capability of the large language model, and uses the Prompt project to construct structured instructions to parse the main text content of the current notes data, forcing the large language model to output a list of customer names and confidence scores in JSON format, and automatically filtering out entities with confidence scores below 0.6.

[0009] Furthermore, the customer consistency verification module adopts a dual-track verification mechanism. The first track performs customer name matching, and the second track performs semantic verification by using a large language model to resolve referential issues. If both tracks fail, it is determined that the core action subject in the main text is inconsistent with the customer in the title.

[0010] Furthermore, the large language model is an open-source model and is deployed privately.

[0011] Furthermore, based on the Kafka message queue module, the data stream of the notes that have passed the quality inspection is synchronized to the knowledge base in real time; when similarity detection is performed again after withdrawal / deletion, the deleted records will no longer be matched.

[0012] Furthermore, knowledge base management, workflow orchestration, and large language model service scheduling are achieved based on the Dify large model development platform.

[0013] Furthermore, the data in the notes includes the account manager's employee ID, the name of the customer visited, the title of the visit schedule, and the visit details.

[0014] Furthermore, the preprocessing includes removing special characters and standardizing full-width and half-width characters.

[0015] Furthermore, the quality inspection report includes risk level, problem type, source tracing of similar segments, and modification suggestions.

[0016] Beneficial effects: 1. Real-time response: End-to-end latency <500ms, supporting real-time early warning; 2. Plagiarism detection rate >95%: 60 percentage points higher than traditional text comparison, supporting semantic-level cross-document plagiarism recognition; 3. Quality inspection efficiency improved by 10 times: The average processing time for a single note is less than 3 seconds, and it supports concurrent submissions from 1000+ account managers; 4. Enhanced management standardization: Issues such as multiple clients and irrelevant information decreased by 90%, and the quality of customer insight data significantly improved; 5. Enhanced interpretability: Quality inspection reports provide AI-powered decision chain traceability, supporting account managers in making targeted improvements. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a collaborative bank customer manager's note-taking quality inspection system according to an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention provides a collaborative bank customer manager's note-taking quality inspection system, including a real-time data acquisition and preprocessing module 1, a knowledge base management module 2, a semantic similarity detection engine 3, a multi-customer intelligent identification module 4, a customer consistency verification module 5, and a quality inspection report generation module 6.

[0020] The real-time data acquisition and preprocessing module 1 provides a standardized input interface to receive and preprocess the notes uploaded from the account manager's terminal. The notes include the account manager's employee ID, the name of the customer visited, the title of the visit schedule, and visit details. Preprocessing involves cleaning the text (removing special characters and standardizing full-width and half-width characters) and then constructing a context for subsequent validation.

[0021] The knowledge base management module 2 is based on the Dify knowledge base. It calls the Embedding model (such as BGE) to vectorize the preprocessed notes data, write it into the vector database (such as Qdrant / Weaviate) to build an index, and build a hierarchical knowledge base according to business lines. It realizes the automatic entry of historical notes data into the database and supports similarity retrieval and access permission isolation.

[0022] The semantic similarity detection engine 3 retrieves Top-K candidate results from the knowledge base using vector similarity, then performs fine-tuning using the Rerank model to comprehensively evaluate semantic relevance. By setting a similarity threshold λ, it identifies semantic plagiarism behavior across account managers and across time dimensions, and outputs semantic plagiarism detection results. The threshold λ ranges from 0.75 to 0.85 and can be set to automatically adjust based on account manager rank and business line. For notes data identified as plagiarized, interference noise (synonym replacement, sentence structure adjustment) can be injected before re-entering the database to increase the difficulty of subsequent plagiarism.

[0023] The multi-customer intelligent identification module 4 defines the customer entity extraction function and parses the preprocessed text content of the notes data, outputting a list of customer names and confidence scores. Specifically, the multi-customer intelligent identification module 4 defines the customer entity extraction function based on the Function Calling capability of the Large Language Model (LLM), and uses the Prompt project to construct structured instructions to parse the text content of the current notes data, forcing the Large Language Model to output a JSON-formatted list of customer names and confidence scores. Entities with a confidence score below 0.6 are automatically filtered. The Large Language Model can be an open-source model, such as Qwen2.5-72B, and can be deployed privately.

[0024] The customer consistency verification module 5 is used to determine whether the core action subject of the preprocessed notes data is consistent with the title "Customer," and outputs the customer consistency detection result. Specifically, the customer consistency verification module 5 adopts a dual-track verification mechanism. The first track performs precise matching of customer names, and the second track performs semantic verification by using a large language model to resolve referential issues. If both tracks fail, it is determined that the core action subject of the text is inconsistent with the title "Customer."

[0025] The quality inspection report generation module 6 is used to aggregate the semantic plagiarism detection results, customer consistency detection results, and customer consistency detection results, and generate a structured quality inspection report. The quality inspection report includes risk level (high / medium / low), problem type, similar fragment source tracing, and modification suggestions.

[0026] It can also synchronize the data stream of the notes that have passed quality inspection to the knowledge base in real time based on the Kafka message queue module 7, with an end-to-end latency of <600 ms and 99.9% of events completing knowledge base synchronization within 1 second; when performing similarity detection after withdrawal / deletion, deleted records will no longer be matched, avoiding "ghost duplicates". Through Kafka transactions and idempotent keys, it achieves 24 / 7 zero duplication and zero loss.

[0027] This system can be deployed on the Dify large model development platform, and can realize knowledge base management, workflow orchestration and large language model service scheduling based on the Dify large model development platform.

[0028] When this system is in operation, it specifically includes the following steps: S1. Receive the Notes data packet, read the customer master data, clean the text (remove special symbols, standardize full and half-width characters), and build a verification context.

[0029] S2. Generate the current embedding vector for the notes, retrieve historical records with similarity exceeding a threshold λ from the knowledge base, and mark them for semantic plagiarism risk; the specific steps are as follows: (1) Call the Embedding service to generate the current note vector V_cur; (2) Perform vector retrieval: Filter historical notes from the same business line, within the last 90 days, and from different account managers in the knowledge base, and return the similarity score set S={s1,s2,...,sk}; (3) Judgment logic: If max(S)>λ, then mark it as "suspected plagiarism" and return the 3 most similar records and the corresponding account manager's employee number; S3. Use the large language model to parse the number of customer entities mentioned in the text. If the number is greater than 1, mark it as a multi-customer violation risk. The Prompt template is as follows: ; S4. Perform customer consistency check to determine whether the core content of the main text matches the customer in the title. If they do not match, mark the customer inconsistency risk.

[0030] S5. A weighted scoring card is used to integrate three dimensions of risk. If the total score is ≥0.6, the quality inspection fails. If the score is 0.3-0.6, it will be subject to manual review. If the score is <0.3, it will be automatically approved and the knowledge base will be updated.

[0031] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A collaborative bank customer manager's note-taking quality inspection system, characterized in that, include: The real-time acquisition and preprocessing module provides a standardized input interface to receive the notes data uploaded by the account manager's terminal and preprocess the received notes data. The knowledge base management module, based on the Dify knowledge base, calls the Embedding model to vectorize and encode the preprocessed notes data, writes it into the vector database to build an index, and builds a hierarchical knowledge base according to business lines to realize the automatic entry of historical notes data into the database, and supports similarity retrieval and access permission isolation. The semantic similarity detection engine is used to retrieve Top-K candidate results from the knowledge base through vector similarity, and then the Rerank model performs fine ranking to comprehensively evaluate semantic relevance; By setting a similarity threshold λ, semantic plagiarism behavior across account managers and across time dimensions is identified, and semantic plagiarism detection results are output. The multi-customer intelligent identification module is used to define customer entity extraction functions, parse the text content of the preprocessed notes data, and output a list of customer names and confidence score detection results. The customer consistency verification module is used to determine whether the core action subject of the preprocessed notes data is consistent with the title "Customer", and outputs the customer consistency detection result. The quality inspection report generation module is used to aggregate the semantic plagiarism detection results, customer consistency detection results, and customer consistency detection results, and generate a structured quality inspection report.

2. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The similarity threshold λ is automatically adjusted based on the account manager's job level and business line.

3. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The multi-customer intelligent identification module defines a customer entity extraction function based on the Function Calling capability of the large language model, and uses the Prompt project to construct structured instructions to parse the main text content of the current notes data, forcing the large language model to output a list of customer names and confidence scores in JSON format, and automatically filtering out entities with confidence scores below 0.

6.

4. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The customer consistency verification module adopts a dual-track verification mechanism. The first track matches the customer name, and the second track performs semantic verification by using a large language model to resolve the referential meaning. If both tracks fail, it is determined that the core action subject in the main text is inconsistent with the customer in the title.

5. The collaborative bank customer manager's note-taking quality inspection system according to claim 4, characterized in that, The large language model is an open-source model, but it is deployed privately.

6. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, Data from notes that have passed quality inspection are synchronized to the knowledge base in real time using the Kafka message queue module; when similarity detection is performed again after a record is withdrawn / deleted, the deleted record will no longer be detected.

7. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, Based on the Dify large model development platform, knowledge base management, workflow orchestration, and large language model service scheduling are implemented.

8. The collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The data in the notes includes the account manager's employee ID, the name of the customer visited, the title of the visit schedule, and the visit details.

9. A collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The preprocessing includes removing special characters and standardizing full-width and half-width characters.

10. A collaborative bank customer manager's note-taking quality inspection system according to claim 1, characterized in that, The quality inspection report includes the risk level, problem type, source tracing of similar segments, and modification suggestions.

Images