Bank active collaborative office method based on large model Agent

Abstract

The invention discloses a bank active collaborative office method based on a large model Agent, and the method comprises the steps: S1, collecting a role-task-tool-data tetrad, and building a domain knowledge base; s2, performing fine tuning by adopting LoRA or QLoRA to obtain a bank field large model; step S3, plug-in RAG + dual Memory is carried out, and an Agent framework is assembled; s4, declaratively registering the tool to a tool warehouse, and obtaining Schema and an authority label; s5, monitoring an external event, and matching a subscription rule to arouse an Agent agent; s6, the intention is distributed to a ReAct template, and a'tool name-input parameter-expected output 'sequence is generated; s7, arranging engine sequence adjustment, permission verification and desensitization, and breakpoint continuous running; and step S8, summarizing and generating deliverable objects, confirming the deliverable objects to the user, and writing a confirmation result into the long-term Memory. A complete closed loop from natural language input to cross-system task automatic execution is realized, and the office efficiency and experience of banking business personnel are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bank digital transformation, and in particular to a bank active collaborative office method based on a large model Agent. BACKGROUND

[0002] With the continuous acceleration of bank digital transformation process, the bank business system is increasingly complex, and business personnel need to frequently operate multiple internal systems (such as CRM, OA, mail system, report platform, etc.) in daily work, and handle a large number of repetitive and rule-based tasks, such as customer information inquiry, meeting arrangement, report generation, to-do list sorting, etc. These operations often involve cross-system data integration, manual judgment and manual input, resulting in low work efficiency, high error rate and poor user experience.

[0003] In this context, intelligent office assistants have become an important tool for banks to improve employee efficiency and optimize business processes. In recent years, office automation systems based on artificial intelligence (AI) technology have gradually emerged, especially dialog-based intelligent assistants centered on large language models (LLM), which have shown strong capabilities in natural language understanding, content generation, and task execution. Some banks have begun to integrate AI assistants into office scenarios, providing functions such as intelligent question answering, email drafting, and schedule reminders.

[0004] Currently, there are mainly three types of solutions for task automatic execution and intelligent collaboration in bank office scenarios: (1) Process automation system based on rule engine or BPM platform This type of system is represented by OA and BPM systems, which drive business processes such as leave, reimbursement, and contract approval through predefined process templates and conditional rules. Users need to submit forms according to fixed processes, and the system only handles state transitions and permission checks.

[0005] Typical products: Pano OA, Zhiyuan Collaboration, K2BPM; technical features: process predefinition, rule hardcoding, no semantic understanding.

[0006] (2) Dialog-based content generation assistant based on large language model Utilize closed-source or open-source large models (such as GPT, Tongyiqiang, Wenxin Yanyan) to provide text question answering and content generation capabilities, which can generate emails, minutes, reports, etc. according to user input, and some products call simple APIs (such as knowledge base search, weather search) through plug-ins.

[0007] Typical products: Microsoft365Copilot, Ali Tongyiqiang, Baidu Wenxin Plug-in; technical features: Prompt-driven, text generation, plug-in tool invocation.

[0008] (3) RPA and AI simple fusion desktop automation script solution The route realizes login, data extraction, report downloading, form filling and other tasks by RPA process robot recording or arranging UI level operations; some manufacturers introduce OCR or NLP as an "entrance" to support selecting robot scripts with natural language.

[0009] Typical products: UiPathAssistant, AutomationAnywhere, and UBots from LeYe; technical features: UI script driven, AI light empowerment, and single-point task automation.

[0010] However, most of the current AI office systems still stay at the single-function level, such as only supporting static question answering or simple task execution, lacking understanding and automatic scheduling capabilities for complex business processes, and unable to truly adapt to the needs of multi-role, multi-task, and multi-system collaboration in the banking scenario. In addition, the existing systems generally have the following problems: Unable to dynamically adjust the interaction strategy according to the user role; Lack of continuous memory and reasoning ability for multi-turn dialogue context; Unable to actively identify business events and automatically invoke services; Unable to automatically decompose natural language instructions into executable task flows; Lack of visualization and controllability of task execution process, and users lack trust in AI behavior.

[0011] Therefore, there is an urgent need for an active collaborative office system for the banking scenario based on large models and agent orchestration, which can integrate large model semantic understanding, task orchestration engine, tool invocation mechanism, and memory system, realize a complete closed loop from natural language input to cross-system task automatic execution, and improve the office efficiency and experience of bank business personnel. SUMMARY

[0012] In view of the above problems, the present application is proposed to provide a bank active collaborative office method based on a large model Agent to overcome the above problems or at least partially solve the above problems.

[0013] According to one aspect of the present application, a bank active collaborative office method based on a large model Agent is provided, which includes: Step S1: Collecting role-task-tool-data quadruples and building a domain knowledge base; Step S2: Fine-tuning a bank domain large model using LoRA or QLoRA; Step S3: Adding RAG+doubleMemory to assemble an Agent framework; Step S4: Declare the tool to the tool warehouse, get the Schema and permission label; Step S5: Listen to external events, match subscription rules to wake up Agent intelligence; Step S6: Intention distribution to ReAct template, generate "tool name-input parameter-expected output" sequence; Step S7: Sequence adjustment of orchestration engine, permission verification + desensitization, breakpoint continuation; Step S8: Summarize the deliverable to user confirmation, and write the confirmation result into long-term memory.

[0014] Optionally, the step S1: collecting role-task-tool-data quadruples, building a domain knowledge base specifically includes: Collecting multi-source heterogeneous data in the bank office field, including business requirement documents, role permission matrix, historical operation logs, external event lists and existing system open interface descriptions; Building a domain knowledge base covering "role-task-tool-data" quadruples, and forming an extensible entity-relation-attribute model for subsequent model training and recall.

[0015] Optionally, the step S2: fine-tuning using LoRA or QLoRA, obtaining a bank domain large model specifically includes: Based on the obtained domain knowledge base, the open source large language model is further pre-trained and fine-tuned to obtain a domain large model with bank domain semantic understanding, task planning and tool calling consciousness; Adopting LoRA or QLoRA low-rank adaptation strategy, retaining general language ability and reducing algorithm cost.

[0016] Optionally, the step S3: external RAG+double memory, assembling Agent framework specifically includes: Taking the domain large model trained in step S2 as the core brain, external RAG recall device, short-term memory cache and long-term memory vector library are assembled into the basic framework of Agent intelligence; Wherein RAG is used for real-time recall of domain knowledge, short-term memory maintains the current session context, and long-term memory deposits user-level preferences and historical execution trajectories.

[0017] Optionally, the step S4: declare the tool to the tool warehouse, obtain the Schema and permission label specifically includes: Register all callable tools in the enterprise to the tool warehouse in a declarative manner, including CRM, OA, email, RPA robot, data API and knowledge base retrieval interface; Each tool generates a natural language description, input-output schema, and permission tags, and writes them into the RAG retrievable knowledge base for Agent dynamic routing.

[0018] Optionally, the step S5 of listening to external events and matching subscription rules to arouse Agent entities specifically includes: Listen to external system change events through the event bus, including new customer signing up, approval node completion, and approaching meeting time, and automatically arouse Agent entities in combination with user-defined subscription rules; Carry event payload and user identity token when aroused, and enter step S6.

[0019] Optionally, the step S6 of distributing intentions to ReAct templates to generate "tool name-input parameter-expected output" sequences specifically includes: After receiving natural language instructions or event arousal signals, the Agent entity first judges the task type, including query, generation, operation, and process type, by the intention distributor; Then, the planner calls the ReAct template to generate an executable step sequence, each step containing a "tool name-input parameter-expected output" triple; The planning result is persisted in the form of a workflow instance to the orchestration engine.

[0020] Optionally, the step S7 of sequential adjustment of the orchestration engine, permission verification + desensitization, and breakpoint continuation specifically includes: The task orchestration engine schedules each step in sequence, automatically matches the optimal tool through the tool router according to the step description, and performs permission verification and data desensitization before calling; Real-time write state, log, and intermediate data to short-term memory and state database during execution, supporting breakpoint continuation and exception rollback.

[0021] Optionally, the step S8 of summarizing and generating deliverables to user confirmation, and writing the confirmation result to long-term memory specifically includes: The content generator summarizes the intermediate results to generate deliverables that meet the bank business specifications; finally, the user confirms through the presentation layer, and the user can one-click modify, re-run, or publish, and write the confirmed result to long-term memory for next personalized recommendation and active arousal.

[0022] Optionally, the deliverables specifically include: customer contact report, meeting minutes, to-do list, and approval summary table.

[0023] The application provides a bank active collaborative office method based on a large model Agent, and the collaborative office method comprises the following steps: S1, collecting a role-task-tool-data quadruple to build a field knowledge base; S2, using LoRA or QLoRA fine tuning to obtain a bank field large model; S3, externally connecting RAG+double memory to assemble an Agent framework; S4, declaratively registering a tool to a tool warehouse to obtain a Schema and a permission tag; S5, listening to external events, matching a subscription rule to an Agent intelligent body; S6, distributing an intention to a ReAct template to generate a "tool name-input parameter-expected output" sequence; S7, sequentially adjusting an arrangement engine, performing permission verification+desensitization, and breaking and continuing running; and S8, summarizing and generating deliverables to user confirmation, and writing a confirmation result into a long-term memory. The bank active collaborative office method based on the large model Agent can integrate semantic understanding of a large model, a task arrangement engine, a tool calling mechanism and a memory system, realize a complete closed loop from natural language input to automatic execution of cross-system tasks, and improve office efficiency and experience of bank business personnel.

[0024] The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application, the content of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the specific embodiments of the application are described below. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical scheme of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description can also be obtained by those skilled in the art without creative labor.

[0026] Figure 1 A flowchart of a bank active collaborative office method based on a large model Agent provided by the embodiments of the application is shown in the figure. Figure 2 A detailed flowchart of a bank active collaborative office method based on a large model Agent provided by the embodiments of the application is shown in the figure. DETAILED DESCRIPTION

[0027] The exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0028] The terms "comprise", "comprising", "include", "including", "have", "having" and any variations thereof in the specification and in the claims of the application and the drawings shall not be construed as excluding any step or element not expressly stated.

[0029] The technical solutions of the application will be further described in detail below in combination with the drawings and examples.

[0030] As shown in the figure, a bank active collaborative office method based on a large model Agent, the collaborative office method comprises the following steps: Figure 1 Step S1: Collecting a role-task-tool-data quadruple, building a domain knowledge base; Step S2: Using LoRA or QLoRA fine-tuning to obtain a bank domain large model; Step S3: External RAG+double Memory, assembling Agent framework; Step S4: Declare the tool to the tool warehouse, obtain Schema and permission tag; Step S5: Listen to external events, match subscription rules to wake up Agent intelligence; Step S6: Intention distribution to ReAct template, generate "tool name-input parameter-expected output" sequence; Step S7: Sequence adjustment of orchestration engine, permission verification+desensitization, breakpoint continuation; Step S8: Summarize the deliverable to user confirmation, and write the confirmation result into long-term Memory.

[0031] As shown in the figure, a bank active collaborative office method based on a large model Agent, the collaborative office method comprises the following steps: Figure 2 Step S1: Collecting a role-task-tool-data quadruple, building a domain knowledge base; Step S2: Using LoRA or QLoRA fine-tuning to obtain a bank domain large model; Step S3: External RAG+double Memory, assembling Agent framework; Step S4: Declare the tool to the tool warehouse, obtain Schema and permission tag; Step S5: Listen to external events, match subscription rules to wake up Agent intelligence; Step S6: Intention distribution to ReAct template, generate "tool name-input parameter-expected output" sequence; Step S7: Sequence adjustment of orchestration engine, permission verification+desensitization, breakpoint continuation; Step S8: Summarize the deliverable to user confirmation, and write the confirmation result into long-term Memory. Step S1: Collecting a role-task-tool-data quadruple, building a domain knowledge base; Step S2: Using LoRA or QLoRA fine-tuning to obtain a bank domain large model; Step S3: External RAG+double Memory, assembling Agent framework; Step S4: Declare the tool to the tool warehouse, obtain Schema and permission tag; Step S5: Listen to external events, match subscription rules to wake up Agent intelligence; Step S6: Intention distribution to ReAct template, generate "tool name-input parameter-expected output" sequence; Step S7: Sequence adjustment of orchestration engine, permission verification+desensitization, breakpoint continuation; Step S8: Summarize the deliverable to user confirmation, and write the confirmation result into long-term Memory.

[0032] Step S2: Based on the domain knowledge base obtained in step S1, the open source large language model is further pre-trained and fine-tuned, and a domain large model with bank domain semantic understanding, task planning and tool calling consciousness is obtained; at the same time, LoRA or QLoRA low-rank adaptation strategy is adopted to retain general language ability and reduce algorithm cost.

[0033] Step S3: Taking the domain large model trained in step S2 as the core brain, external RAG recall device, short-term Memory cache and long-term Memory vector library are assembled into an Agent intelligence basic framework; wherein RAG is used for real-time recall of domain knowledge, short-term Memory maintains the current session context, and long-term Memory deposits user-level preferences and historical execution trajectories.

[0034] Step S4: Register all callable tools within the enterprise to the tool warehouse in a declarative manner, including CRM, OA, email, RPA robots, data API, and knowledge base retrieval interface; generate natural language description, input and output Schema, and permission tags for each tool, and write them into the RAG retrievable knowledge base for Agent dynamic routing recall.

[0035] Step S5: Listen to external system change events (such as new customer signing, approval node completion, meeting time approaching, etc.) through the event bus, combine with user-defined subscription rules, and automatically wake up Agent intelligent agents; wake up with event payload and user identity token, enter step S6.

[0036] Step S6: After receiving natural language instructions or event wake-up signals, the Agent intelligent agent first judges the task type (query, generation, operation, process type) by the intent distributor, then calls the planner to generate executable step sequences based on the ReAct template, each step containing a "tool name-input parameter-expected output" triple; The planning result is persisted in the form of a workflow instance to the orchestration engine.

[0037] Step S7: The task orchestration engine schedules each step in order, automatically matches the optimal tool through the tool router according to the step description, and performs permission verification and data desensitization before calling; Real-time write state, log and intermediate data to short-term Memory and state database during execution, support breakpoint resume and exception rollback.

[0038] Step S8: After all steps are executed, the content generator summarizes the intermediate results to generate deliverables (such as customer contact reports, meeting minutes, to-do lists, and approval summary tables) that meet bank business specifications; Finally, return to the user for confirmation through the presentation layer, the user can modify, rerun or publish with one key, and write the confirmed results to the long-term Memory for next personalized recommendation and active wake-up.

[0039] A bank active collaborative office method based on large model Agent, comprising: In step S1, collect multi-source heterogeneous data in the bank office field, including business requirement documents, role permission matrix, historical operation logs, external event lists, and existing system open interface descriptions, build a domain knowledge base covering "role-task-tool-data" four-tuple, and form an extensible entity-relation-attribute model for subsequent model training and recall; Training set data sample: ``` {Instruction: This is a bank office task generation task, please generate a professional executable task plan according to the given information, Input: Role: Public customer manager; Requirement: Query A company's account flow for the past three months and generate a fund analysis report; Permission: Access to account system and report system; Event: None, Output: Task plan: 1. Call the account system API to get the flow for the past three months; 2. Call the report system to generate a fund analysis chart; 3. Merge the chart and text to generate a PDF report; 4. Send the report to the customer manager's email; 5. Record the task completion status in the user memory bank. ``` In step S2, the large model is pre-trained and fine-tuned on the domain knowledge base to obtain a domain large model with semantic understanding, task planning, and tool calling awareness in the banking field; using LoRA low-rank adaptation strategy, the general language ability is preserved and the algorithm cost is reduced, and a professional bank office domain large model is obtained.

[0040] In step S3, the domain large model trained in step S2 is used as the core brain, and RAG recallers, short-term Memory caches, and long-term Memory vector libraries are attached to assemble an Agent intelligent agent basic framework; RAG is used to recall domain knowledge in real time, short-term Memory maintains the current session context, and long-term Memory deposits user-level preferences and historical execution trajectories, which are used to generate high-professional and accurate task plans.

[0041] In step S4, all callable tools in the enterprise are registered in the tool warehouse in a declarative manner, including CRM, OA, email, RPA robot, data API, and knowledge base retrieval interface; natural language description, input and output Schema, and permission tags are generated for each tool, and written into the knowledge base retrievable by RAG for Agent dynamic routing; knowledge base example: ``` {Tool name: account flow query API, Description: Query enterprise customer's specified month account flow, Input Schema: {customerId, startDate, endDate}, Output Schema: {transactionList}, Permission tag: public customer manager} ``` In step S5, the external system change events (such as new customer signing up, approval node completion, meeting time approaching, etc.) are monitored through the event bus, and the user's pre-defined subscription rules are combined to automatically arouse the Agent intelligent agent; when aroused, it carries the event payload and user identity token, and enters step S6.

[0042] In step S6, the Agent intelligent agent disassembles the task disassembly rule elements to form a generated process through prompts, which is used for step-by-step generation; prompt words are as follows: ``` {Please generate an executable task plan based on the following information: Role: [Role], Requirements: [Requirement Description], Permissions: [Accessible System List], Events: [Event Payload].

[0043] Requirements: 1. Disassemble into no more than 5 executable steps; 2. Each step contains tool name, input parameter, and expected output; 3. Follow the permission tags and do not exceed the authority; 4. Output in JSON format.

[0044] } ``` In step S7, the Agent intelligent agent combines the tool description knowledge injected in step S4 with the large model trained in step S2 to generate each element of the task plan in turn: tool selection, input parameter, expected output, and exception handling strategy, and caches intermediate results through short-term Memory to support breakpoint continuation and exception rollback.

[0045] In step S8, the Agent intelligent agent submits the task plan generated by the process to the task orchestration engine, which calls the corresponding tools in sequence and writes the state and log in real time; after all steps are executed, the content generator summarizes the intermediate results to generate a deliverable that meets the bank's business specifications (such as customer contact report, meeting minutes, to-do list, and approval summary table); finally, the presentation layer returns to the user for confirmation, and the user can modify, rerun, or publish with one click, and the confirmed results are written into long-term Memory for next personalized recommendation and active arousal; a new task plan example is as follows: ``` {Task Name: Generate A Company's Capital Analysis Report, Tool Sequence: 1. Account Stream Query API → Obtain Transaction Data, 2. Report Generation API → Generate Chart, 3. Document merge API -> output PDF, 4. Email sending API -> push user, 5. Memory write API -> deposit preferences, Expected result: user mailbox receives PDF report, task status is "complete", and memory bank is updated with query preferences. Beneficial effects: 1. Reduce process expansion cost and improve response flexibility Existing rule engines / BPM systems hardcode processes in templates, and any changes require re-development, testing, and online, resulting in high expansion costs and delayed response. The invention uses a natural language-driven dynamic orchestration engine to analyze user intent in real time and generate process paths that match roles, permissions, and scenarios, eliminating the need for hard-coded maintenance, significantly reducing expansion costs and improving business response speed.

[0046] 2. Achieve natural language to cross-system closed-loop automation Large pure model dialogue assistants only output text suggestions and lack the ability to call business systems, making it impossible to form a closed loop. The invention adds a planner and workflow engine outside the model to automatically break down natural language instructions into executable steps and call corresponding tool interfaces in sequence to complete cross-system data reading and writing, approval, notification, and other operations, achieving complete automation from "text" to "result".

[0047] 3. Establish sustainable memory and proactive service capabilities Existing systems have no cross-session memory and cannot actively sense external events, resulting in repeated clarification and passive waiting for each interaction. The invention introduces a user-level vector memory bank and Profile persistence layer to deposit personal preferences, commonly used instructions, and permission ranges. At the same time, through an event bus and subscription rule pool, it actively listens to external system changes and automatically invokes agents once conditions are triggered, forming an "event-driven" proactive service capability, reducing the number of manual triggers and improving user experience.

[0048] The above detailed description further explains the purpose, technical solutions, and beneficial effects of the invention. It should be understood that the above is only a specific implementation of the invention and does not limit the scope of protection. Any modifications, equivalent replacements, improvements, etc. within the spirit and principles of the invention should be included in the scope of protection.

Claims

1. A proactive collaborative office method for banks based on a large-scale agent model, characterized in that, The collaborative office methods include: Step S1: Collect the role-task-tool-data quadruple and build a domain knowledge base; Step S2: Fine-tune using LoRA or QLoRA to obtain a large-scale model for the banking sector; Step S3: Add RAG+ dual memory to assemble the Agent framework; Step S4: Declaratively register the tool to the tool repository to obtain the schema and permission tags; Step S5: Listen for external events and match the subscription rules to activate the Agent. Step S6: Distribute the intent to the ReAct template to generate the "tool name-input parameters-expected output" sequence; Step S7: Adjust the orchestration engine order, verify permissions and de-identify data, and resume execution from breakpoints; Step S8: Summarize the generated deliverables and confirm them with the user, then write the confirmation results into long-term memory.

2. The method for proactive collaborative work in banks based on a large-scale agent model as described in claim 1, characterized in that, Step S1: Collecting the role-task-tool-data quadruple and building a domain knowledge base specifically includes: Collect multi-source heterogeneous data from the banking office sector, including business requirement documents, role and permission matrices, historical operation logs, external event lists, and descriptions of existing system open interfaces; Construct a domain knowledge base covering the "role-task-tool-data" quadruple and form an extensible entity-relationship-attribute model for subsequent model training and retrieval.

3. The method for proactive collaborative work in banks based on a large-scale agent model as described in claim 1, characterized in that, Step S2: Fine-tuning using LoRA or QLoRA to obtain a large-scale model for the banking domain specifically includes: Based on the obtained domain knowledge base, the open-source large language model is further pre-trained and fine-tuned with instructions to obtain a domain large model with semantic understanding, task planning and tool calling awareness in the banking field; We adopt a low-rank adaptation strategy of LoRA or QLoRA to retain general language capabilities and reduce computing power costs.

4. The method for proactive collaborative work in banks based on a large-scale agent model as described in claim 1, characterized in that, Step S3: assembling the Agent framework by attaching external RAG+ dual memory specifically includes: Using the domain-specific large model trained in step S2 as the core brain, and attaching RAG recaller, short-term memory cache and long-term memory vector library, the basic framework of Agent is assembled. RAG is used to retrieve domain knowledge in real time, short-term memory maintains the current session context, and long-term memory stores user-level preferences and historical execution trajectories.

5. The method for proactive collaborative work in banks based on a large-scale agent model according to claim 1, characterized in that, Step S4: Declaratively registering the tool to the tool repository and obtaining the schema and permission tags specifically includes: Register all callable tools within the enterprise to the tool repository in a declarative manner, including CRM, OA, email, RPA robots, data APIs, and knowledge base retrieval interfaces; Each tool generates a natural language description, input / output schema, and permission tags, and writes them into the RAG's searchable knowledge base for recall during agent dynamic routing.

6. The method for proactive collaborative work in banks based on a large-scale agent model according to claim 1, characterized in that, Step S5: Listening for external events and matching subscription rules to activate the Agent intelligent body specifically includes: By monitoring external system change events through the event bus, including new customer signings, approval node completions, and approaching meeting times, and combining these with user-defined subscription rules, the Agent intelligent agent is automatically activated. Upon activation, carry the event payload and user identity token, and proceed to step S6.

7. The method for proactive collaborative work in banks based on a large-scale agent model as described in claim 1, characterized in that, Step S6: Distributing the intent to the ReAct template and generating the "tool name-input parameters-expected output" sequence specifically includes: After the agent receives a natural language instruction or event-triggered signal, the intent dispatcher first determines the task type, including query, generation, operation, and process types. The planner is then invoked to generate an executable sequence of steps based on the ReAct template, with each step containing a triplet of "tool name - input parameters - expected output"; The planning results are persisted to the orchestration engine as workflow instances.

8. The method for proactive collaborative work in banks based on a large-scale agent model according to claim 1, characterized in that, Step S7: Orchestration engine order adjustment, permission verification + data masking, and breakpoint resume specifically include: The task orchestration engine schedules each step sequentially, and the tool router automatically matches the optimal tool based on the step description, performing permission verification and data anonymization before calling the tool. During execution, the status, logs, and intermediate data are written back to short-term memory and the status database in real time, supporting breakpoint resumption and exception rollback.

9. A method for proactive collaborative work in banks based on a large-scale agent model, as described in claim 1, is characterized in that... Step S8: Summarizing the generated deliverables and confirming them with the user, and writing the confirmation results into long-term memory, specifically includes: The content generator summarizes intermediate results and generates deliverables that conform to banking business standards. Finally, it returns confirmation to the user through the presentation layer. The user can modify, rerun, or publish with one click, and the confirmed result is written to long-term memory for future personalized recommendations and proactive reminders.

10. A bank proactive collaborative office method based on a large model agent according to claim 9, characterized in that, The deliverables specifically include: customer contact reports, meeting minutes, to-do lists, and approval summary tables.

Images