A Method and System for Managing Institutional Documents Based on Multi-Agent Collaboration

Abstract

A method and system for managing institutional documents based on multi-agent collaboration are disclosed. The method includes: performing semantic matching in a repository based on user input questions to obtain recall information; organizing the recall information and user input questions into prompts and providing them to multiple agents; a central agent assigning tasks to a first agent, a second agent, and a third agent; the first agent performing step-by-step reasoning on the user input questions through guided statements and generating intermediate reasoning steps, and providing a first result; the second agent analyzing the user input questions in conjunction with the prompts and providing a second result; the third agent, acting as a fine-tuning model in the financial domain, performing financial domain analysis on the user input questions and outputting a third result; and the central agent forming a final result based on the first to third results, and triggering subsequent business processes based on the final result, thereby achieving intelligent analysis of diverse customer feedback content and increasing the accuracy of retrieval.

Description

Technical Field

[0001] This application relates to the field of combining artificial intelligence with finance, and in particular to a method and system for managing institutional documents based on multi-agent collaboration. Background Technology

[0002] Financial institutions possess a vast number of policy documents that are frequently updated and often exist in multimodal formats, including text, scanned copies, and audio / video training materials. Traditional manual retrieval, comparison, and compliance review methods are inefficient, prone to omissions, and fail to meet the real-time regulatory requirements. In recent years, Large Language Modeling (LLM) technology has been introduced into policy management and has achieved some success. However, general-purpose LLMs lack sufficient differentiation in vertical scenarios within the financial sector, and the output results lack validation, leading to compliance risks. Summary of the Invention

[0003] The purpose of this disclosure is to provide a method and system for managing institutional documents based on multi-agent collaboration, so as to solve the problems existing in the prior art.

[0004] According to a first aspect of the present disclosure, a method for managing institutional documents based on multi-agent collaboration is provided, comprising:

[0005] Semantic matching is performed on the repository based on the user's input question to obtain recall information;

[0006] The recall information and the user input question are organized into a prompt message and provided to a multi-agent system, which includes a central agent, a first agent, a second agent, and a third agent.

[0007] The central agent assigns tasks to the first agent, the second agent, and the third agent.

[0008] The first intelligent agent performs step-by-step reasoning on the user input question through guided statements and generates intermediate reasoning steps, and then feeds back the first result;

[0009] The second intelligent agent, acting as a general large language model, analyzes the user input question by combining the prompt information and provides a second result.

[0010] The third intelligent agent, as a fine-tuning model in the banking field, performs banking-related analysis on the user input question and outputs a third result;

[0011] The central intelligent agent forms a final result based on the first result to the third result, and triggers subsequent business processes based on the final result.

[0012] In some embodiments, the system document management method further includes: performing multiple rounds of execution on the user input question, evaluating the first to the third results of the multiple rounds, and taking the result with the most occurrences of the answer as the final result.

[0013] In some embodiments, performing semantic matching in the repository based on user-input questions to obtain recall information includes:

[0014] Vector matching calculation is performed between the vector representation of the user input question and the vector representation of multiple small text blocks in the repository to obtain multiple small text blocks;

[0015] Based on the obtained multiple small text blocks, multiple large text blocks are obtained in the repository; and

[0016] The obtained multiple small text blocks and multiple large text blocks are organized into the recall information.

[0017] In some embodiments, it also includes:

[0018] The policy document is parsed and divided into small text blocks using a parsing method corresponding to the various file formats of the policy document;

[0019] Combine multiple small text blocks into a large text block;

[0020] Mapping small text blocks to text vectors; and

[0021] The repository stores the large text block, the small text block, the policy document, the association between the large text block and the small text block, and the association between the small text block and the text vector.

[0022] In some embodiments, the user input question and the text, images, and video frames in the banking policy document are mapped to vector representations in a shared semantic space through a multimodal embedding model to achieve cross-modal semantic alignment.

[0023] In some embodiments, triggering subsequent business processes includes:

[0024] The final result is converted into an MCP service call according to the Model Control Protocol format to drive the execution of nodes in the compliance review, risk identification, or approval process.

[0025] In some embodiments, converting the final result into an MCP service call according to the Model Control Protocol format includes:

[0026] The calling parameters of the MCP service include business type identifier, risk level, compliance conclusion and required approval node. The business system automatically matches and starts the corresponding MCP service call according to the calling parameters to drive the execution of nodes in the compliance review, risk identification or approval process, and sends the execution status back to the business system after the process node is completed.

[0027] According to a second aspect of the present disclosure, a system document management system based on multi-agent collaboration is provided, comprising:

[0028] The data processing module is used to divide banking system documents into multiple small text blocks, combine the small text blocks into large text blocks, vectorize the text blocks, and store the small text blocks, the large text blocks, the vector representations of the small text blocks, the relationships between the large text blocks and the small text blocks, and the relationships between the small text blocks and the vector representations in the repository.

[0029] The retrieval module is used to perform semantic matching in the repository based on user-input questions in order to obtain recall information;

[0030] A multi-agent system, comprising a central agent, a first agent, a second agent, and a third agent.

[0031] The central agent assigns tasks to the first agent, the second agent, and the third agent.

[0032] The first intelligent agent performs step-by-step reasoning on the user input question through guided statements and generates intermediate reasoning steps, and then feeds back the first result;

[0033] The second intelligent agent, acting as a general large language model, analyzes the user input question by combining the prompt information and provides a second result.

[0034] The third intelligent agent, as a fine-tuning model in the banking field, performs banking-related analysis on the user input question and outputs a third result;

[0035] The central intelligent agent forms a final result based on the first result to the third result, and triggers subsequent business processes based on the final result.

[0036] In some embodiments, the data processing module includes:

[0037] Cross-modal semantic alignment is achieved by mapping the user input questions and the text, images, and video frames in the policy documents to vector representations in a shared semantic space through a multimodal embedding model.

[0038] In some embodiments, the first agent to the third agent perform multiple rounds of analysis and feedback on the user input question, the central agent evaluates the first to the third results in multiple rounds, and takes the result with the most occurrences as the final result.

[0039] In some embodiments, the method further includes: configuring a business triggering process.

[0040] According to a third aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described institutional document management method.

[0041] The solution provided in this disclosure coordinates three distinct agents (first agent: large model and thought chain; second agent: semantic matching and enhanced retrieval; third agent: fine-tuning model in the financial field) through a central agent, enabling intelligent analysis of diverse customer feedback and increasing retrieval accuracy. Furthermore, this solution utilizes MCP (Model Control Protocol) and function call mechanisms to transmit the final results in a structured manner to the business system or middleware process service, triggering backend process execution, thereby achieving intelligent analysis of diverse customer feedback and enhancing retrieval accuracy. Attached Figure Description

[0042] The above and other objects, features and advantages of the present disclosure will become clearer from the following description of embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0043] Figure 1 This is a block diagram of institutional document management based on multi-agent collaboration provided in an embodiment of this disclosure;

[0044] Figure 2A and Figure 2B This is a flowchart of the preprocessing of banking system documents provided in this embodiment of the disclosure;

[0045] Figure 3 The following is given Figure 2A and Figure 2B A flowchart showing how the preprocessed document is divided into multiple small text blocks. Detailed Implementation

[0046] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements are indicated by similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale. Furthermore, certain well-known parts may not be shown.

[0047] The following description of embodiments of the present disclosure is based on examples, but the embodiments of the present disclosure are not limited to these embodiments. In the detailed description of the embodiments of the present disclosure below, certain specific details are described in detail. Those skilled in the art can fully understand the embodiments of the present disclosure without these details. To avoid obscuring the essence of the embodiments of the present disclosure, well-known methods, processes, flows, elements, and circuits are not described in detail.

[0048] Unless the context explicitly requires it, the terms "comprising," "including," and similar terms throughout the specification and claims should be interpreted as encompassing rather than exclusive or exhaustive; that is, meaning "including but not limited to." In the description of embodiments of this disclosure, it should be understood that terms such as "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more.

[0049] Large Language Models (LLMs) refer to deep learning language models with hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. They can be widely used in fields such as Natural Language Processing (NLP) and Computer Vision. However, when they are actually applied to a specific task, such as Visual Question Answering (VQA), Image Captioning (IC), Text Summarization, and Machine Translation, it is usually necessary to fine-tune the model using samples related to that specific task.

[0050] In this paper, an intelligent agent refers to a computational entity that performs calculations on external inputs and provides outputs based on its internal logic. The internal logic can be a simple internal bank rule, a complex learning model, or a combination of both.

[0051] To address the issue of managing institutional documents in financial institutions, this application proposes an institutional document management system based on multi-agent collaboration, such as... Figure 1 As shown, the system mainly consists of four functional layers: data layer 110, retrieval layer 120, control layer 130, and multi-agent layer 140.

[0052] The data layer 110 processes the input banking policy documents and stores relevant information. Its main modules include a data processing module 101 and a storage module 102. The data processing module 101 parses the policy documents using parsing methods corresponding to various file formats, then segments them into small text blocks 1011 using a segmentation model. These small text blocks 1011 are then merged into a large text block 1012. Furthermore, a multimodal embedding model 1013 maps the small text blocks 1011 into text vectors. Size limits for the small text blocks 1011 and large text blocks 1012 can be predefined; for example, small text blocks 1011 can be between 200 and 400 characters, and large text blocks 1012 between 800 and 1200 characters. Based on these size limits, the segmentation model adaptively divides paragraphs into small text blocks, incorporating semantic understanding. For example, the segmentation model places the main clause and several subordinate clauses within multiple small text blocks, and then merges these small text blocks into a single large text block based on the semantic relationships between them, achieving semantic integrity. During model training, the size limits for large and small text blocks can be adjusted based on the model's performance.

[0053] The repository 102 includes a file repository 1021, a relational database 1022, and a vector database 1023. The file repository 1021 stores banking policy documents in various file formats. The relational database 1022 stores the relationships between small text blocks 1011 and large text blocks 1012. The vector database 1023 stores the text vectors of small text blocks 1011 and the relationships between small text blocks 1011 and their text vectors. The file types of the banking policy documents include, but are not limited to, Word, PPT, PDF, audio, video, and images, as shown in the diagram. These content formats can also be organized within the same banking policy document. The parsing methods corresponding to the file formats include, but are not limited to, image recognition (OCR), mathematical formula recognition, table recognition, and layout recognition, as shown in the diagram. The multimodal embedding model 1013 can map text, images, and video frames in the banking policy documents to vector representations in a shared semantic space to achieve cross-modal semantic alignment. For audio and video, they can be directly mapped to vector representations, or they can be transcribed into text content through speech recognition, then the text content is divided into multiple small text blocks, and multiple audio and video segments are segmented accordingly. Then, the vector representations of each small text block are associated with the multiple audio and video segments in the repository, for example, by using the CLIP model for mapping.

[0054] Figure 2A and Figure 2B This is a flowchart of the preprocessing of banking system documents provided in this embodiment.

[0055] exist Figure 2A and Figure 2B In the process, steps S201 to S208 involve reading data into memory. For Word files, they need to be converted into PDF files. For PDF files, if the file is particularly large, the offset parameter controls the offset during reading, and the length limit parameter controls the length of reading. These two parameters are used to read the data in batches. If the file is small, it is read in binary form and loaded directly into memory.

[0056] Steps S209 to S221 further determine whether the PDF file is an image-based PDF (scanned document) or a text-based PDF. The specific judgment logic is as follows: If most pages (more than half) of the main content in the PDF consists of a large image (the image occupies more than 50% of the page), then the PDF is a scanned document and needs OCR processing; otherwise, the PDF is considered a text-based PDF. Then, a portion of pages is randomly selected, and if any page contains more than 100 characters of text, the PDF is considered a text-based PDF. Then, the average number of characters per page in the PDF is calculated. If the average number of characters per page exceeds 100, the PDF is considered text-based. Next, it's checked if the PDF is a special type of scanned document characterized by having no valid content after removing duplicate images. If not, it's considered a text-based PDF. Then, it's checked if more than half of the pages have a vast majority of images (at least 5 images representing more than 80% of the total images on that page) that are very elongated (i.e., width exceeding 90% of the page width and at least 4 times the height, or height exceeding 90% of the page height and at least 4 times the width). If this proportion of pages is less than 50%, the PDF is considered text-based in this dimension. During the scanning of the PDF content, invalid or unreadable characters are detected. If not, it's classified as text-based. Finally, considering all six conditions, if all six conditions are met, the file is classified as a text-based PDF; if any condition is not met, the file is classified as an image-based PDF.

[0057] Steps S222 to S224 refer to specifying the running device and device mode before the actual operation. The device modes are cpu, cuda, and npu. If cuda or npu is selected, the video memory capacity will be automatically calculated and the batch ratio will be adjusted adaptively.

[0058] Steps S225 to S232 process PDFs. Image-based PDFs require an additional OCR processing step compared to text-based PDFs. Other steps include layout analysis, formula processing, and table processing. OCR refers to recognizing the content of an image region as text. Different models are needed for different languages, and the configuration file specifies the appropriate models for each language. Formula processing consists of two stages: formula detection and formula recognition. Formula detection identifies the location of the formula, while formula recognition identifies the specific content within the formula region and converts the mathematical expression image into LaTeX. Table recognition comprises two parts: table text recognition and table structure recognition. Table text recognition identifies the text content within the table, while table structure recognition, by default, identifies the table structure.

[0059] Figure 3 The following is given Figure 2A and Figure 2B The flowchart shows how the preprocessed output document is divided into multiple small text blocks. Specifically, the document layout is identified in the following order: receiving the pre-identified file S301, identifying chapters S302, identifying appendices S303, identifying main headings S304, identifying entries S305, and identifying the main text S306. Then, the content is cut into segments S307 according to different business line document characteristics strategies. Figure 3 The flowchart shown can also represent the operations within a large language model. To enable a large language model to better segment text according to business logic in the financial field, model fine-tuning is necessary. This involves obtaining a training dataset by labeling features in banking policy documents. During training, the model parameters are gradually adjusted based on the set hyperparameters and optimization algorithms to enable the model to recognize the labeled content. The effectiveness of the fine-tuned model is then evaluated on a validation dataset to ensure it can accurately segment text based on key features, resulting in small text blocks of the smallest semantic unit.

[0060] Back Figure 1The retrieval layer 120 is used to retrieve relevant text blocks from the repository 102 based on the user input question 122, and organize the relevant text blocks into prompt information and pass it to the central agent 132. The user input question 122 can be speech or text, which is converted into a question vector 123 through a multimodal embedding model, or the speech can be converted into text and then into a question vector 123 through a dedicated speech program. Then, the question vector 123 and the vector representations of multiple small text blocks in the vector database 1023 are used for vector matching calculation to obtain multiple small text blocks. Then, multiple large text blocks associated with the obtained small text blocks are obtained in the relational database 1022. Finally, the obtained multiple small text blocks and multiple large text blocks are organized into recall information 124. Then, the user input question 122 and the recall information 124 are organized into prompt information 131 and passed to the central agent 132. The user input question 122 can come from the input interface of the business system. Vector matching algorithms used for vector matching calculations can be cosine similarity, Euclidean distance, or Jaccard similarity, with cosine similarity being a preferred algorithm. Furthermore, vector matching calculations can be performed in batches using the vector database 1023 to achieve high execution efficiency.

[0061] The control layer 130 coordinates three distinct agents (A: large model and thought chain; B: semantic matching and RAG; C: fine-tuning model in the financial field) through the central agent 132, achieving accurate feedback on diverse user input questions. Based on a deep understanding and analysis of the user input questions, and a precise grasp of each agent (A, B, C), the central agent 132 assigns tasks to agents A, B, and C, and provides different task information.

[0062] The technology of agent A primarily relies on API calls to a large language model and the application of Chain-of-Thought (CoT) technology. Its core lies in constructing guiding statements designed to facilitate effective reasoning within the large language model. First, a suitable large language model needs to be selected. Then, the guiding statements explicitly require the model to "think step by step" or provide examples of reasoning steps. For example, the guiding statement given to agent A is: "You are a banking business analysis expert. Please carefully analyze the following customer feedback, and think about and answer the following questions step by step: 1. What is the customer's core demand? 2. What key events or issues are described in the feedback? 3. What banking business processes or products might these issues involve?" After receiving such a guiding statement, the large language model will generate intermediate inference steps according to the guidance and finally output structured label information, i.e., the first result. To improve the stability and controllability of the output, other techniques can be combined, such as controlling the randomness of the generated text, using stop sequences to ensure that the output format meets expectations, and post-processing the output of the large language model (such as extracting key information and converting it to a standard label format). Agent B further analyzes the user input question based on the prompt information 131 organized from the user input question 122 and recall information 124, and provides the second result.

[0063] Agent C is a fine-tuned version of a large language model in the financial field. It performs financial analysis on user input questions and outputs a third result.

[0064] The central intelligent agent 132 receives the first to third results from agents A to C to form the final result 143, and triggers subsequent business processes based on the final result. Triggering subsequent business processes includes converting the final result into an MCP service call to drive the execution of nodes in compliance review, risk identification, or approval processes. Specifically, the central intelligent agent 132 uses the MCP (Model Control Protocol) and function call mechanism to transmit the final result in a structured manner to the bank's business system or middleware process service. For example, when the system identifies that a certain item needs to enter the compliance review process, it packages the intent and semantic parameters (e.g., business type identifier, risk level, compliance conclusion, required approval node) through the MCP protocol and transmits them to the review process service node. The review process service node automatically matches and initiates the corresponding process node, such as risk identification or manual review, without manual intervention, to drive the execution of nodes in the compliance review, risk identification, or approval process, and returns the execution status to the business system after the process node is completed. This mechanism achieves the goals of seamless process integration, closed-loop business linkage, and intelligent automatic execution, significantly improving the efficiency and compliance of banking procedures.

[0065] Furthermore, Figure 1 The system shown may also include a process configuration module, which provides a visual configuration interface that allows for flexible definition of process strategies based on system type, approval level, and business scenario. Banking personnel can configure different trigger conditions, notification rules, and approval paths to achieve personalized customization and auditable execution of business processes. The process strategy configuration ensures that the system can adapt to different banking business needs and facilitates subsequent expansion.

[0066] Furthermore, Figure 1 The system shown may also include: a central agent 132 controlling the execution of the user input question in multiple rounds, with each round of execution obtaining the first to third results through agents A to C, and then the evaluation module 142 evaluating the first to third results of the multiple rounds, counting the first to third results of the multiple rounds, and taking the result with the most occurrences as the final result.

[0067] It should be noted that during the initial system deployment, due to insufficient basic data, the feedback results of agent A are primarily used. As the amount of basic data increases but still falls short of the ideal level, the feedback results of agents A and B are primarily used. Once the amount of basic data is sufficiently large, the feedback results of agents A through C are employed. Furthermore, in practice, concurrent processes (or threads) are used to improve the execution efficiency when using agents A through C simultaneously.

[0068] Accordingly, this application also provides a method for managing institutional documents corresponding to the aforementioned institutional document management system. This method involves the following steps.

[0069] In step S10, semantic matching is performed in the repository based on the user-input question to obtain recall information.

[0070] In step S20, a prompt message composed of user input questions and recall information is provided to multiple agents, including a central agent, a first agent, a second agent, and a third agent. The central agent assigns tasks to the first, second, and third agents and provides task information. The first agent performs stepwise reasoning on the user input questions through guided statements and generates intermediate reasoning steps, and provides a first result. The second agent analyzes the user input questions in conjunction with the prompt message and provides a second result. The third agent, as a fine-tuning model in the financial field, performs financial field analysis on the user input questions and outputs a third result. Finally, the central agent forms a final result based on the first to third results and triggers subsequent business processes based on the final result.

[0071] It should be noted that the system document management method based on multi-agent collaboration corresponds to the system document management system mentioned above, so the system document management method will be introduced in a relatively brief manner here.

[0072] In practical implementation, the methods and systems mentioned above can be executed by computer devices, which can be terminals or servers. This includes the systems mentioned above, the execution of the involved models, and model training, all of which can be performed by terminals and / or servers. For example, the server can perform the relevant processes for model fine-tuning, and the trained model can be deployed to the server. During model application, the terminal can obtain user input questions and send them to the server; the server can then invoke various services to jointly execute and obtain the question results.

[0073] In a specific implementation, the above-mentioned scheme can be formed into a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the various operations in the above embodiments.

[0074] The embodiments of this disclosure are as described above. These embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made based on the above description. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the embodiments of this disclosure, thereby enabling those skilled in the art to make good use of the embodiments of this disclosure and modifications based on them. The embodiments of this disclosure are limited only by the claims and their full scope and equivalents.

Claims

1. A method for managing institutional documents based on multi-agent collaboration, characterized in that, include: Semantic matching is performed in the repository based on user-input questions to obtain recall information; The recall information and the user input questions are organized into prompts and provided to a multi-agent system. The multi-agent system includes a central agent, a first agent, a second agent, and a third agent. The first agent is a large model and thought chain; the second agent is semantic matching and RAG (Research and Analysis Group); and the third agent is a fine-tuning model for the financial domain. The central intelligent agent assigns tasks to the first intelligent agent, the second intelligent agent, and the third intelligent agent, and provides different task information to each of them; The first intelligent agent guides the selected large language model to perform step-by-step reasoning on the user input question and generate intermediate reasoning steps through guiding statements, and then feeds back the first result; The second intelligent agent analyzes the user input question in conjunction with the prompt information and provides a second result; The third intelligent agent performs financial analysis on the user input question and outputs a third result; The central intelligent agent activates the first intelligent agent, the second intelligent agent, and the third intelligent agent, and forms a final result based on the results of the three activated intelligent agents. The final result is then converted into an MCP service call to drive the execution of nodes in the compliance review, risk identification, or approval process of financial institutions. The three activated intelligent agents perform multiple rounds of parallel analysis and feedback on the user input questions. The step of performing semantic matching in the repository based on user input questions to obtain recall information includes: Vector matching calculation is performed between the vector representation of the user input question and the vector representation of multiple small text blocks in the repository to obtain multiple small text blocks; Based on the multiple small text blocks obtained, multiple large text blocks are obtained in the repository; The obtained multiple small text blocks and multiple large text blocks are organized into the recall information.

2. The method for managing institutional documents according to claim 1, characterized in that, The central agent evaluates the first to the third results in multiple rounds, and takes the result with the most occurrences as the final result.

3. The method for managing institutional documents according to claim 1, characterized in that, Also includes: The policy document is parsed and divided into small text blocks using a parsing method corresponding to the document's file format. Combine multiple small text blocks into a large text block; Map small text blocks to text vectors; as well as The repository stores the large text block, the small text block, the policy document, the association between the large text block and the small text block, and the association between the small text block and the text vector.

4. The method for managing institutional documents according to claim 1 or 3, characterized in that, Cross-modal semantic alignment is achieved by mapping the user input questions and the text, images, and video frames in the policy documents to vector representations in a shared semantic space through a multimodal embedding model.

5. The method for managing institutional documents according to claim 1, characterized in that, The final result is converted into an MCP service call according to the Model Control Protocol format, including: The calling parameters of the MCP service include the business type identifier, risk level, compliance conclusion, and required approval node. The business system matches and initiates the corresponding MCP service call according to the calling parameters to drive the execution of nodes in the compliance review, risk identification, or approval process, and returns the execution status to the business system after the process node is completed.

6. A system document management system based on multi-agent collaboration, characterized in that, include: The data processing module is used to divide the policy document into multiple small text blocks, combine the small text blocks into a large text block, vectorize the text blocks, and store the small text blocks, the large text blocks, the vector representations of the small text blocks, the relationships between the large text blocks and the small text blocks, and the relationships between the small text blocks and the vector representations in the repository. The retrieval module is used to perform vector matching calculations on the vector representation of the user input question and the vector representation of multiple small text blocks in the repository to obtain multiple small text blocks, obtain multiple large text blocks in the repository based on the obtained multiple small text blocks, and organize the obtained multiple small text blocks and multiple large text blocks into recall information. The multi-agent system includes a central agent, a first agent, a second agent, and a third agent. The first agent is a large model and a thought chain, the second agent is semantic matching and RAG, and the third agent is a fine-tuning model in the financial field. The central agent assigns tasks to the first agent, the second agent, and the third agent. The first intelligent agent guides the selected large language model to perform step-by-step reasoning on the user input question and generate intermediate reasoning steps through guiding statements, and then feeds back the first result; The second intelligent agent analyzes the user input question by combining the recall information and the user input question with the prompt information, and feeds back a second result; The third intelligent agent performs banking-related analysis on the user input question and outputs a third result; The central intelligent agent activates the first intelligent agent, the second intelligent agent, and the third intelligent agent, and forms a final result based on the results of the three activated intelligent agents. The final result is then converted into an MCP service call to drive the execution of nodes in the compliance review, risk identification, or approval process of financial institutions. The three activated intelligent agents perform multiple rounds of parallel analysis and feedback on the user input questions.

7. The institutional document management system according to claim 6, wherein the data processing module comprises: Cross-modal semantic alignment is achieved by mapping the user input questions and the text, images, and video frames in the policy documents to vector representations in a shared semantic space through a multimodal embedding model.

8. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the institutional document management method as described in any one of claims 1 to 5.

Images