Intelligent valuation and accounting system and method based on AI large model technology

Abstract

The application discloses an intelligent valuation accounting system and method based on AI large model technology, and comprises the following steps: receiving a natural language instruction input by a user, determining a target Agent through a double-path matching mechanism of historical dialogue library similarity retrieval and large model reasoning; calling a financial business process rule library, decomposing the user instruction into a sub-task chain containing document uploading, information extraction, interface calling and data verification; analyzing the financial document uploaded by the user, extracting product basic information, cost structure and account information key fields by using large model entity recognition and relationship extraction technology; mapping the extracted key fields to valuation system interface parameters, generating a calling request with a unique operation serial number and recording operation logs; performing abnormal hierarchical processing: interrupting the task and requesting manual intervention when core field conflicts or missing are detected, and delaying the processing of non-core field abnormalities.

Description

Technical Field

[0001] This invention relates to the field of financial asset management technology, specifically to an intelligent valuation and accounting system and method based on AI large-scale model technology. Background Technology

[0002] In the field of financial asset management, traditional valuation and accounting systems have significant shortcomings: Operational inefficiency: Scenarios such as the establishment and tiering of fund products require manual step-by-step maintenance of system settings (such as product parameters and valuation schemes), involving independent operation of multiple functional modules; for example, creating a new account set requires sequentially entering basic product information, fee structure, account information, etc., which takes a long time for a single operation; Document processing shortcomings: The extraction of key information from financial documents (contracts, custody agreements) relies on manual labor and cannot automatically link cross-document data (e.g., after extracting the fee rate from the product contract, it is necessary to manually compare the custody agreement account information). Inflexible exception handling: When data conflicts or missing data occur, the system can only interrupt the process and issue a full alert, lacking a hierarchical handling mechanism (such as treating missing core fields and warnings for non-critical fields the same), leading to frequent manual intervention. Fragmented system interaction: Users need to switch between different functional modules and cannot drive end-to-end task chains through natural language commands. Summary of the Invention

[0003] This invention aims to at least partially solve one of the technical problems in related technologies. To this end, one objective of this invention is to propose an intelligent valuation and accounting method based on AI large-scale model technology, comprising: It receives natural language commands input by the user and determines the target Agent through a dual-path matching mechanism of similarity retrieval from the historical dialogue database and inference from a large model. The financial business process rule base is invoked to break down user instructions into a chain of sub-tasks, including document upload, information extraction, API call, and data verification. Parse user-uploaded financial documents and use large model entity recognition and relationship extraction technologies to extract key fields of basic product information, fee structure, and account information; The extracted key fields are mapped to the valuation system interface parameters, a call request with a unique operation serial number is generated, and the operation log is recorded. Execution of anomaly classification: When a core field conflict or missing value is detected, the task is interrupted and manual intervention is requested; non-core field anomalies are handled later. Generate a visual summary report, outputting task execution status, exception handling trajectory, and a list of pending items.

[0004] Preferably, the dual-path matching mechanism includes: The Agents whose similarity to the current user's command exceeds a preset threshold are retrieved from the historical dialogue database and selected as the first candidates. All Agent description information is input into the large model for inference matching as the second candidate; Prioritize selecting agents with consistent matching results from two different paths; otherwise, select candidate agents with higher similarity.

[0005] Preferably, the information extraction step includes: When key fields are missing, data is completed across multiple uploaded documents according to preset priority; It supports users to define custom field extraction rules and dynamically add new rules to the financial business process rule library.

[0006] Preferably, the anomaly classification processing specifically includes: Level 1 anomaly: Product name or account information conflict or missing, forcibly interrupt the task and pop up a manual confirmation window; Level 2 anomaly: Rate format error or value exceeds limit, allowing users to choose immediate processing or delayed processing; Level 3 Exception: Non-critical field warning, logs exception information but does not interrupt the task flow.

[0007] Preferably, the interface call step includes: Perform rule and logic validation on the extracted fields to ensure that the data conforms to the valuation system interface specifications; A user token authentication mechanism is used to verify the identity of the call request. Automatic retry when API call times out; the number of retries is configurable.

[0008] Preferably, the visual summary report is a structured Excel file, containing: The completed operation list specifies the operation name, the fields involved, and the execution result. Exception handling records, noting conflicting fields, handling methods, and operation times; The pending items are categorized and displayed in groups based on missing data and rule conflicts.

[0009] An intelligent valuation and accounting system based on AI large-scale model technology includes: S1. Intent recognition module, used to parse natural language commands and match the target Agent; S2. Multi-Agent Collaboration Engine: Performs task decomposition, document parsing, and API calls; S3. Dynamic rule base, storing financial business process rules and custom extraction rules; S4. Audit feedback module, which generates operation logs and visual summary reports.

[0010] Preferably, the multi-agent collaboration engine includes: The task decomposer breaks down sub-task chains based on a financial business process rule base. The document parser utilizes the entity recognition and relationship extraction capabilities of large model systems to extract key fields. The exception handler implements a three-level exception classification strategy and initiates a human-machine collaboration window.

[0011] Preferably, the dynamic rule base supports: When a new financial document type is added, the field extraction logic is extended; When regulatory policies change, business rules are automatically updated.

[0012] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent valuation and accounting method based on AI large model technology.

[0013] The above-described solution of the present invention has at least the following beneficial effects: First, natural language command parsing and task triggering: receiving user natural language commands (such as "create a new account set based on reference combination A"), and determining the target Agent through a dual-path matching mechanism: Search the historical dialogue database, match agents with a similarity greater than 90%, input all agent descriptions into the large model for inference and matching, prioritize the consistent results from both paths, otherwise take the agent with the highest similarity; Secondly, task decomposition and intelligent document parsing involve calling the financial business process rule base to break down instructions into sub-task chains: (This process is repeated three times in the original text.) Upload product contracts / custody agreements and other documents, use large model entity recognition to extract fields such as product name and management fee rate, extract the "custody fee - fee rate value" logic through relation extraction, and fill in missing fields across documents (such as extracting the bank of account opening from the custody agreement). Next, system calls and exceptions work together to map fields to valuation system interface parameters (e.g., product code → C_PORT_CODE), and corresponding level three exception handling is executed, for example: The trigger condition for a Level 1 exception is a missing product name or account. The handling method is to interrupt the task and force manual confirmation via a pop-up window. The triggering conditions for a Level 2 exception are exceeding the rate limit or a format error, and the handling method is to prompt the user to choose immediate or delayed processing; The trigger condition for a Level 3 exception is a warning in a non-critical field, and the handling method is to log the exception but not interrupt the task flow. Finally, the results are fed back and audited, generating a structured Excel report, which includes: completed operations: such as "[Account Set Creation] Successfully copied 2 data records", abnormal events: such as "Management fee rate conflict: 1.2% → 1.5% (manual overwrite)", and pending items: a list categorized by data missing and rule conflict. The management system of this invention significantly improves operational efficiency, with natural language-driven automation replacing manual operations and reducing product setup time. Furthermore, it features intelligent anomaly handling, a three-level anomaly classification strategy to reduce manual intervention, and automatic repair of conflicting fields. It also provides full-process traceability, with a unique operation serial number and structured logs supporting second-level problem node location, improving compliance audit efficiency. The system also boasts enhanced compatibility, with a custom rule base supporting dynamic expansion, allowing adaptation to new financial document types without system reconstruction.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0016] Figure 1 This is a flowchart of an intelligent valuation and accounting method based on AI large model technology provided in an embodiment of the present invention; Figure 2 This is a functional architecture diagram of the asset management platform provided in this embodiment of the invention; Figure 3 This is a system relationship diagram provided in the asset management platform in this embodiment of the invention; Figure 4 This is a flowchart of an intelligent valuation and accounting system based on AI large model technology provided in an embodiment of the present invention; Figure 5 This is a flowchart of the media deployment and execution process provided in this embodiment of the invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "circumferential," and "radial," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0023] The intelligent valuation and accounting method based on AI large model technology according to embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Please see Figure 1 The intelligent valuation and accounting method of this invention specifically includes the following process: First, the system receives natural language instructions input by the user and determines the target Agent through a dual-path matching mechanism. One path retrieves Agents with similarity exceeding a preset threshold from the historical dialogue database as the first candidate, while the other path inputs all Agent description information into a large model for inference and matching as the second candidate. Agents with consistent matching results from both paths are prioritized; otherwise, the candidate Agent with higher similarity is selected. Second, the system calls the financial business process rule base to decompose the user instruction into a sub-task chain including document upload, information extraction, interface call, and data verification. Then, the system parses the financial documents uploaded by the user and uses large model entity recognition and relationship extraction technology to extract key fields of basic product information, fee structure, and account information. When key fields are missing, data is supplemented from multiple uploaded documents according to a preset priority. Next, the system maps the extracted key fields to the valuation system interface parameters, generates a call request with a unique operation serial number, and records the operation log. Before the interface call, the extracted fields are validated by rules and logic, and a Token authentication mechanism is used for identity authorization verification. When the call times out, it automatically retryes the request a configurable number of times. During the execution of the above steps, the system synchronously performs hierarchical exception handling: when a core field conflict or missing value is detected, the task is forcibly interrupted and a manual confirmation window pops up for Level 1 exception handling; when a rate format error or value exceeding the limit is detected, the user can choose to handle it immediately or delay it as Level 2 exception handling; when a non-critical field warning is detected, only the exception information is recorded without interrupting the task flow, as Level 3 exception handling. Finally, the system generates a visual summary report, outputting the task execution status, exception handling trajectory, and a list of pending items.

[0025] First, natural language command parsing and task triggering: receiving user natural language commands (such as "create a new account set based on reference combination A"), and determining the target Agent through a dual-path matching mechanism: Search the historical dialogue database, match agents with a similarity greater than 90%, input all agent descriptions into the large model for inference and matching, prioritize the consistent results from both paths, otherwise take the agent with the highest similarity; Secondly, task decomposition and intelligent document parsing involve calling the financial business process rule base to break down instructions into sub-task chains: (This process is repeated three times in the original text.) Upload product contracts / custody agreements and other documents, use large model entity recognition to extract fields such as product name and management fee rate, extract the "custody fee - fee rate value" logic through relation extraction, and fill in missing fields across documents (such as extracting the bank of account opening from the custody agreement). Next, system calls and exceptions are coordinated to map fields to valuation system interface parameters (e.g., product code → C_PORT_CODE), and corresponding level three exception handling is executed, for example: The trigger condition for a Level 1 exception is a missing product name or account. The handling method is to interrupt the task and force manual confirmation via a pop-up window. The triggering conditions for a Level 2 exception are exceeding the rate limit or a format error, and the handling method is to prompt the user to choose immediate or delayed processing; The trigger condition for a Level 3 exception is a warning in a non-critical field, and the handling method is to log the exception but not interrupt the task flow. Finally, the results are fed back and audited, generating a structured Excel report, which includes: completed operations: such as "[Account Set Creation] Successfully copied 2 data records", abnormal events: such as "Management fee rate conflict: 1.2% → 1.5% (manual overwrite)", and pending items: a list categorized by data missing and rule conflict. The management system of this invention significantly improves operational efficiency, with natural language-driven automation replacing manual operations and reducing product setup time. Furthermore, it features intelligent anomaly handling, a three-level anomaly classification strategy to reduce manual intervention, and automatic repair of conflicting fields. It also provides full-process traceability, with a unique operation serial number and structured logs supporting second-level problem node location, improving compliance audit efficiency. The system also boasts enhanced compatibility, with a custom rule base supporting dynamic expansion, allowing adaptation to new financial document types without system reconstruction.

[0026] In this embodiment, the dual-path matching mechanism includes: retrieving Agents from the historical dialogue database whose similarity to the current user's command exceeds a preset threshold (e.g., 90%) as the first candidate; inputting all Agent description information into a large model for inference matching as the second candidate; prioritizing the selection of Agents whose matching results are consistent across both paths, otherwise selecting the candidate Agent with higher similarity. Scenario Description: The user initiates the task of "creating a new account set using reference combination Alpha" via natural language commands, and the system executes a dual-path matching mechanism to determine the target Agent.

[0027] I. Historical Dialogue Database Search: The system calculates the similarity between user commands and records in the historical dialogue database; For example, the historical instruction "Reference combination Beta create product account set" was found to have a 92% similarity to the current instruction (exceeding the 90% threshold), and was initially matched to "Product creation Agent".

[0028] II. Large-scale model inference matching: At the same time, the system will input the description information of all agents (such as "the product establishment agent is responsible for handling the task of creating a new account set for the fund product") into the large model; The large model analyzes the semantics of user commands and infers that the "product establishment agent" needs to be called.

[0029] III. Arbitration of results from both paths: If the two matching results are consistent (both are "Product established Agent"), the system will directly trigger the workflow for that Agent.

[0030] Alternative scenario: If the highest similarity in the historical dialogue database is only 85% (below 90%), and the large model inference result is "product established Agent", the system will prioritize the large model matching result (because model inference takes priority when the similarity does not meet the standard).

[0031] Technical effect: The dual-path redundancy matching mechanism significantly improves the agent's positioning accuracy, avoids the risk of mismatches from a single search path, and ensures accurate response to complex natural language commands.

[0032] In this embodiment, the information extraction steps include: when key fields are missing, data is completed across multiple uploaded documents according to a preset priority; user-defined field extraction rules are supported, and new rules are dynamically added to the financial business process rule library; Scenario description: When processing the establishment of a fund product, users need to extract complete information from multiple documents to create an account set.

[0033] I. Key Field Extraction and Missing Field Identification: The user uploads a PDF document titled "XX Fund Product Contract". The system extracts fields such as product name, product type, and management fee rate using large model entity recognition technology, but fails to extract complete custodian bank account information (e.g., the specific branch name of the opening bank is missing).

[0034] II. Cross-document auto-completion mechanism: The system automatically identifies missing account information fields and, based on preset document priority rules (e.g., the custody agreement has higher priority than the product contract), retrieves and extracts the missing custody bank account details (including the opening bank name, account number, and branch information) from the uploaded "Custody Agreement" document to complete the field completion.

[0035] III. Custom Rule Extension Applications: When processing new structured fund contracts, users found that the system failed to recognize the "performance fee accrual benchmark" field unique to the document. Users can customize and add extraction rules using the rule templates provided by the system: map the synonyms of the "performance compensation accrual benchmark" field in the document to the "performance fee benchmark" field built into the system, and dynamically add the rule to the financial business process rule library; When the system processes similar documents in the future, it can automatically identify and extract this unique field.

[0036] Technical benefits: Through the cross-document completion mechanism, the system ensures the integrity and accuracy of key information, avoiding business process interruptions due to missing information in a single document; through the custom rule extension mechanism, the system can flexibly adapt to new document formats and business needs, significantly improving the system's adaptability and scalability.

[0037] In this embodiment, the anomaly classification processing is specifically as follows: Level 1 anomaly: Product name or account information conflict or missing, forcibly interrupt the task and pop up a manual confirmation window; Level 2 anomaly: Rate format error or value exceeds limit, allowing users to choose immediate processing or delayed processing; Level 3 Exception: Warning for non-critical fields, logs exception information but does not interrupt the task flow; Scenario Description: During the execution of the "Create a new account set based on reference combination Gamma" task, different levels of anomalies were triggered during the data verification process.

[0038] I. Triggering a Level 1 Exception (Critical): During system data verification, it was discovered that the product name "XX Growth Hybrid Securities Investment Fund" extracted from the product contract was completely identical to a product name already existing in the valuation system, causing a uniqueness conflict. Handling method: The system immediately interrupts all subsequent operations in the current task chain and pops up a forced confirmation window in the dialog interface, prompting the user: "Duplicate product name detected. Please confirm whether to continue creating or modifying the product name." The task process is paused, waiting for the user to make a manual decision and input.

[0039] II. Triggering a Level 2 Exception (Medium): The system checked the management fee rate extracted from the document and found that its format was "one hundred and five percent", which does not conform to the numerical format "1.5" required by the system interface. Handling method: The system will display a pop-up message on the interface: "Management rate format error, please select a handling method" and provide users with two options: "Correct Now" and "Process Later". Users can choose to manually enter the correct format immediately, or choose to process later to complete other tasks first. The system will record this error for later notification.

[0040] III. Triggering a Level 3 Exception (Minor): The system discovered that the investment manager's name field extracted from the document had not yet been entered into the associated professional information database. Handling method: The system will not interrupt the task, nor will it pop up a window to force the user to handle it immediately; it will record a warning message in the background operation log: "Investment manager information has not been filed in the system". At the same time, this message will also be included in the "Remarks" or "Warning" column of the final generated visual summary report, prompting the user to handle it as needed in subsequent processes.

[0041] Technical effect: Through a three-level anomaly classification and handling mechanism, the system can intelligently distinguish the severity of problems and take differentiated response strategies. This ensures that core business processes are not blocked by non-critical issues, while strictly controlling problems that may cause serious errors, achieving the best balance between processing efficiency and risk control.

[0042] In this embodiment, the interface call steps include: performing rule and logic validation on the extracted fields to ensure that the data conforms to the valuation system interface specification; using a user token authentication mechanism to verify the identity authorization of the call request; and automatically retrying when the interface call times out, with the number of retries being configurable. Scenario Description: The system has completed the extraction of key fields from the product creation document and is ready to call the valuation and accounting system's interface to create a new set of accounts.

[0043] I. Parameter Mapping and Generation: The system converts the extracted element information into parameters required by the valuation system interface according to the preset interface mapping rules; for example, it maps the "product name" extracted from natural language to the interface parameter C_PORT_NAME; maps the "product code" to C_PORT_CODE; and maps the "establishment date" to D_BUILD.

[0044] II. Parameter Verification: Before encapsulating the request, the system calls the built-in business rule library to validate the parameters.

[0045] Rule validation: Check whether the product code conforms to the naming convention of "6 digits".

[0046] Logical verification: Check whether the management fee rate is within the industry standard range of 0.1% to 5.0%.

[0047] If the system finds that the product code format is "AB1234" (containing letters), which does not conform to the rules, the system will immediately interrupt the call process and prompt the user in the dialog interface: "Product code format is incorrect, please change it to 6 digits".

[0048] III. Request Encapsulation and Security Authentication: The validated parameters are encapsulated into a structured JSON request body according to the valuation system interface specification.

[0049] The system automatically adds the current user's authentication token to the HTTP request header for the valuation system to verify authorization and ensure that the call is legitimate and authorized.

[0050] IV. Interface Calls and Exception Handling: The system sends a POST request to the / api / v1 / port / create endpoint of the valuation system.

[0051] Timeout Handling: The system has a preset timeout period of 3 seconds. If no response is received within 3 seconds, the system will automatically retry. The number of retries can be adjusted according to the operation and maintenance configuration (e.g., a maximum of 2 retries).

[0052] Error handling: If the valuation system returns an error response code (e.g., 400 means "parameter error"), the system will parse the error information, categorize it, and inform the user in a user-friendly manner (e.g., "API call failed: product code already exists, please replace it"). At the same time, the complete error log will be recorded in the operation log for technical personnel to troubleshoot.

[0053] Technical effects: Through rigorous parameter validation, secure authentication mechanisms, and structured exception handling processes, the accuracy of data and system security for each interface call are ensured; the automated retry mechanism effectively copes with temporary network or system fluctuations, improving the final execution success rate of tasks and the robustness of the system.

[0054] In this embodiment, the visual summary report is a structured Excel file, which includes: a list of completed operations, listing the operation name, fields involved and execution results; an exception handling record, marking conflicting fields, handling methods and operation times; and a category table of pending items, grouped and displayed according to data missing and rule conflicts. Scene Description After a fund company completed the creation of the "New Energy Industry Fund" account set, the system generated a visual Excel report with the file name: "New Energy Industry Fund_Account Set Creation Report_20250806.xlsx"; Report Structure Analysis Worksheet 1: List of Completed Operations

[0055] Worksheet 2: Exception Handling Track

[0056] Worksheet 3: Category Table of Items to be Processed

[0057] Dynamic optimization: Since there are no level 3 anomalies, the "level 3 anomaly record table" is automatically hidden.

[0058] Technical process implementation I. Data Collection: Real-time capture of logs from each node in the task chain: "[2025-08-06 14:20:33] API call: / create-account?code=NY2025 Response code 200" [2025-08-06 14:22:01] Anomaly: Management fee rate conflict (1.5% vs 1.2%).

[0059] II. Template Filling: Use a preset Excel template (including formulas and conditional formatting): Automatic status coloring: Success → Green, Warning → Yellow, Failure → Red; Field association: "Funds Flow Setting Warning" is associated with the pending items worksheet.

[0060] Dynamic cropping: No "Level 3 Anomaly" records detected → Delete the corresponding worksheet (to save users reading burden).

[0061] Technical benefits: This visual summary report integrates fragmented information that was originally scattered in logs, databases, and interface prompts into a standardized document with a clear structure and easy-to-understand presentation. It provides users with a panoramic view of task execution, greatly facilitating business review, compliance checks, and audit trails, and realizing a closed loop of process management.

[0062] like Figures 2-4As shown, the intelligent valuation and accounting system based on AI large model technology includes: an intent recognition module, used to parse natural language instructions and match target agents; a multi-agent collaboration engine, which performs task decomposition, document parsing and interface calls; a dynamic rule base, which stores financial business process rules and custom extraction rules; and an audit feedback module, which generates operation logs and visual summary reports. The intelligent valuation and accounting system of this invention adopts a layered architecture, including an intent recognition module, a multi-Agent collaboration engine, a dynamic rule base, and an audit feedback module. The intent recognition module, located at the system front end, is responsible for receiving natural language commands input by the user (including text or text converted from speech), and outputs a uniquely identified target agent identifier through a built-in dual-path matching mechanism, serving as a signal to trigger downstream tasks. The multi-Agent collaboration engine is the central control unit of the system, integrating three core capabilities: task decomposition, document parsing, and exception handling (its internal structure is described in [reference needed]). Figure 3 The engine receives the target Agent identifier and reads the corresponding business process rules from the dynamic rule base, responsible for the orchestration, scheduling, and status management of the entire valuation and accounting task. The dynamic rule base independently stores financial business process rules, field extraction rules, and interface mapping rules, and supports online dynamic expansion—when new financial document types are added or regulatory policies change, the rules can be updated through the rule management interface without restarting the system. The audit feedback module, located in the system backend, receives operation logs (each log entry corresponds to a unique operation serial number), exception handling traces, and the final task status reported by the multi-Agent collaboration engine in real time. Internally, it includes a log storage unit and a report generator, which can output structured Excel reports or standardized log interfaces for external audit systems to call as needed. Furthermore, the system interacts with external valuation and accounting systems, financial systems, etc., through a unified gateway; before calling external interfaces, the multi-Agent collaboration engine must undergo token authentication and parameter verification by the audit feedback module, and the call results are synchronously written back to the logs. The arrows between the modules clearly indicate the data flow: user commands are passed sequentially through the intent recognition module to the multi-agent collaboration engine, the engine reads rules from the dynamic rule base, calls external system interfaces, and finally writes the operation record into the audit feedback module.

[0063] The multi-agent collaboration engine comprises three sub-components: a task decomposer, a document parser, and an exception handler. The task decomposer breaks down user instructions into a chain of dependent sub-tasks based on standard business process rules obtained from a dynamic rule base (e.g., the "product creation" rule chain includes four sub-tasks: document upload, information extraction, API call, and data verification), and schedules them for execution one by one.

[0064] The task decomposer receives the matched Agent identifier and user instructions, retrieves the standard business process rules corresponding to the Agent from the dynamic rule base (for example, the "product establishment" rule chain includes four sub-tasks: document upload, information extraction, interface call, and data verification), generates a sub-task chain with dependencies based on the rules, and dispatches the sub-tasks one by one to the document parser or directly calls the external interface.

[0065] The document parser is responsible for executing the document parsing and field extraction steps in the sub-task chain. Internally, it encapsulates the entity recognition model and relation extraction model of the large model. The document parser receives financial documents uploaded by users (such as PDF, Word, etc.) and outputs a structured set of key fields (including product name, management fee rate, custody account information, etc.). When a field is found to be missing in the current document, the document parser completes the data from other uploaded documents according to the preset document priority (e.g., custody agreement has higher priority than product contract). If a custom field that the system cannot recognize is encountered, an exception handler is triggered to start the human-machine collaboration window.

[0066] The exception handler monitors the execution status of the task chain throughout its operation. Internally, it maintains a three-tiered exception judgment rule table (Level 1: Core field conflicts or missing values; Level 2: Rate format errors or value exceeding limits; Level 3: Non-critical field warnings). Based on the exception level, the exception handler performs three processing operations: For Level 1 exceptions, it interrupts the task chain and displays a manual confirmation window, waiting for the user to input corrected data; for Level 2 exceptions, it displays a prompt allowing the user to choose immediate or delayed processing; for Level 3 exceptions, it only records the warning information to the operation log without interrupting the task flow. The exception handler also manages a human-machine collaboration window, through which it interacts with the user, receives manually corrected data, and sends the processing results back to the task decomposer to restore or continue the task chain execution.

[0067] Solid arrows represent direct calls between the task decomposer and the document parser, while dashed arrows represent monitoring and interruption control of both by the exception handler. The dynamic rule base is read by both the task decomposer and the exception handler. The multi-agent collaboration engine consists of three well-defined, collaborative sub-components, ensuring the programmability of complex financial transactions and the flexibility of exception handling.

[0068] The document parser is responsible for executing the document parsing and field extraction steps in the sub-task chain. Internally, it calls the entity recognition model and relation extraction model of the larger model to convert unstructured financial documents (PDF, Word, etc.) into a structured set of key fields (such as product name, management fee rate, custodian account information). When a field is missing in the current document, the document parser completes it from other uploaded documents according to preset document priorities. The exception handler monitors the execution status of the task chain throughout. Internally, it maintains three levels of exception judgment rules, and performs three operations according to the exception level: "interrupt the task and pop up a manual confirmation window", "allow the user to choose immediate or delayed processing", and "log only without interrupting the process". The exception handler also manages the human-machine collaboration window, through which it interacts with the user, receives manually corrected data, and sends the results back to the task decomposer to resume execution. The task decomposer and document parser are directly called by solid arrows, while the exception handler uses dashed arrows to indicate the monitoring and interruption control relationship between the task decomposer and document parser. The dynamic rule base is read by both the task decomposer and the exception handler. The multi-agent collaboration engine is not a single module, but consists of three sub-components with distinct responsibilities that cooperate with each other, thus ensuring the programmability of complex financial transactions and the flexibility of exception handling.

[0069] The system execution flow of this invention begins with the system receiving natural language commands input by the user. It then enters the Agent matching stage, where the system performs parallel similarity retrieval from the historical dialogue database and large-scale model inference matching. The final target Agent is determined based on the consistency of the two results. After determining the target Agent, the system calls the financial business process rule base, decomposing the command into sub-tasks such as document upload, information extraction, interface call, and data verification. In the document parsing stage, the system uses a large-scale model to extract key fields from financial documents. If conflicts or missing core fields (such as product name or account information) are detected, the process transitions to a first-level exception handling branch—interrupting the task and displaying a manual confirmation window for user processing. If a rate format error or numerical limit exceeding is detected, the process transitions to a second-level exception handling branch—allowing the user to choose immediate correction or delayed processing. If a warning is detected in a non-critical field, the process transitions to a third-level exception handling branch—logging only without interrupting the process. Verified data is mapped to valuation system interface parameters, and after token authentication, external interfaces are called and operation logs are recorded. If an interface call times out, the system automatically retryes the call a configurable number of times. Finally, the system selectively generates a visual summary report containing a list of completed operations, an exception handling trajectory, and a classification table of pending items, based on whether there are any pending items or abnormal records.

[0070] Scenario: An insurance asset management company deploys the ChatFA system. System architecture and data flow (e.g.) Figure 4 (As shown).

[0071] S1. Intent recognition module in operation Input: User voice command "Create a new account set for retirement target 2045"; deal with: 1. Voice-to-text: "Create a new account set for the pension goal of 2045"; II. Key Entities in the Large Model Analysis: Product Name: Pension Target 2045, Operation Type: New Account Set; Output: Matching product establishes Agent (confidence level 99%).

[0072] S2. Multi-Agent Collaboration Engine Execution I. The task decomposer calls the dynamic rule base: Retrieve the "Insurance Product Establishment" rule chain → Generate subtasks: 1. Upload the "Product Contract" and "Actuarial Report" 2. Extraction: Product type, investment ratio, guaranteed rate of return 3. Call the valuation system interface.

[0073] II. Document parser execution Extracting fields from "Actuarial Report.pdf": Guaranteed rate of return: 2.8% (based on large model relationship extraction) ESG investment ratio: 35% (new custom field added)

[0074] III. Exception Handling Response: The system detected that there was no corresponding interface parameter for "ESG investment ratio" → triggering a level 2 exception. A pop-up window appears: "A new field has been detected. Please map it to the valuation system."

[0075] S3. Dynamic rule base update Triggering condition: The regulator issues the "Guidelines for ESG Disclosure in Insurance Asset Management"; Automatic update process: Key words from the new regulatory rules: "ESG investment ratio" and "carbon neutral portfolio"; Generate rule update package: "{"Field name": "ESG investment ratio", "Interface Mapping": "ENV_RATE", "Validation rule": "Value ≥ 30% and ≤ 100%" Synchronize to all Agents → Effective time: 5 minutes (traditional systems require system shutdown for upgrade).

[0076] S4. Audit Feedback Module Output Input data: Operation log: The product's Agent called the API 3 times; Error Log: ESG field mapping error - resolved.

[0077] Generate report:

[0078] Report content: "Operation completed: [Interface call] ENV_RATE=35%" Pending items: None (the anomaly has been closed).

[0079] In this embodiment, the multi-Agent collaboration engine includes: a task decomposer, which breaks down sub-task chains based on a financial business process rule base; a document parser, which uses the entity recognition and relationship extraction capabilities of a large model to extract key fields; and an exception handler, which implements a three-level exception classification strategy and initiates a human-machine collaboration window. Scenario: A user initiates a task to "handle the liquidation of the fund product Omega" via natural language commands. The system completes this complex business process through the collaborative components of the multi-Agent collaboration engine.

[0080] I. Task decomposer operation: Once the target Agent is determined, the task decomposer is immediately started; it accesses the scenario rules related to "product liquidation" in the financial business process rule base.

[0081] Based on the standardized processes defined in the rule base, the task decomposer breaks down the complex requirement of "product liquidation" into an ordered chain of executable subtasks with dependencies, for example: Subtask 1: Upload the liquidation audit report and legal opinion.

[0082] Subtask 2: Extract key fields such as liquidation date, remaining asset allocation plan, and tax treatment instructions from the uploaded documents.

[0083] Subtask 3: Call the asset liquidation interface of the valuation and accounting system.

[0084] Subtask 4: Call the final payment interface of the financial management system.

[0085] Subtask 5: Perform liquidation data verification to ensure all operations comply with regulatory requirements.

[0086] II. Document parser operation: When the workflow reaches the "upload document" subtask, the document parser is invoked.

[0087] It receives liquidation audit reports in PDF format uploaded by users, and uses the entity recognition capabilities of large models to accurately locate and extract entity fields such as "liquidation date" and "final net value".

[0088] Subsequently, by leveraging the relation extraction capabilities of the large model, the logical relationship between "remaining assets" and "shareholder distribution ratio" is understood and extracted, transforming unstructured text information into structured data that the system can process.

[0089] III. Exception Handler Operation: In the "Perform liquidation data verification" subtask, the exception handler is activated.

[0090] It discovered that the extracted "liquidation date" was earlier than the latest transaction date in the system, which violated the business logic (Level 3 exception: logical error class). The exception handler logged this event as a warning but did not interrupt the process.

[0091] Subsequently, when the asset liquidation interface was called, the interface returned the error "Account status is abnormal, liquidation cannot be completed" (Level 1 exception: severe). The exception handler immediately interrupted the entire task chain, froze all subsequent subtasks, and popped up a dialog box to prompt the user: "The target account status is abnormal, the liquidation process has been suspended, please prioritize the account issue."

[0092] After resolving the account issue, users can choose to resume the task from the point of interruption. At this point, the exception handler detected a blurry stamp in an auxiliary document, which may affect the audit (Level 2 exception: Medium). Provide users with the following options: "Document defects detected. Please select [Re-upload Now] or [Continue Process]".

[0093] Technical effects: Through the division of labor and cooperation of three specialized components—task decomposer, document parser, and exception handler—the system can efficiently and reliably handle complex and strictly regulated financial transactions such as "product liquidation." The entire process is highly automated, and necessary human-machine collaboration nodes are embedded through a hierarchical exception handling mechanism, ensuring business compliance and operational security.

[0094] In this embodiment, the dynamic rule base supports: expanding field extraction logic when a new financial document type is added; and automatically triggering business rule updates when regulatory policies change. Scenario: During system operation, it needs to cope with the emergence of new financial documents and changes in external regulatory policies. The dynamic rule base plays a core supporting role in this process.

[0095] I. Expanded support for new document types: A fund management company launched an innovative product whose legal documents (such as the "Sustainable Investment Disclosure Report") include a new key field, "Green Industry Investment Ratio"; the initial document parsing rule base of the system did not define the extraction logic for this field.

[0096] Users (such as business experts) can customize a new extraction rule based on the structure and content characteristics of the new document through the rule management interface provided by the system. The rule specifies the field name ("Green Industry Investment Ratio"), its location characteristics in the document (such as the chapter title and keyword environment), and the internal system parameters to which it should be mapped.

[0097] After the user completes the rule definition and confirmation, the system dynamically adds the new rule to the financial business process rule base; thereafter, when the system processes similar documents, it can automatically identify and extract the unique field without requiring program code-level upgrades or refactoring.

[0098] II. Responding to changes in regulatory policies: Financial regulators have issued new accounting standards that require adjustments to the valuation methods for certain financial assets (e.g., from amortized cost to fair value measurement). This change directly affects the rules governing the “valuation treatment” scenario in valuation accounting.

[0099] The system's built-in dynamic update mechanism is automatically triggered; this mechanism identifies the policy adjustment by accessing external regulatory information sources or by being manually activated by the administrator.

[0100] Based on the requirements of the new guidelines, the system automatically evaluates the existing business rules in the rule base that involve asset valuation methods and prompts business experts to review and update them.

[0101] Business experts optimized and adjusted the specific operational steps and verification rules for the "valuation processing" scenario through the system interface, based on the new regulatory requirements; the updated rules took effect immediately, ensuring that all subsequent valuation and accounting tasks automatically complied with the latest regulatory compliance requirements.

[0102] Technical benefits: The dynamic rule base's expansion and update mechanism gives the system strong adaptability and evolution capabilities; it can quickly respond to business innovations (new documents, new fields) and changes in the external environment (new policies) through configuration rather than coding, greatly reducing system maintenance costs, extending its life cycle, and continuously ensuring the compliance and accuracy of business processing.

[0103] A computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements an intelligent valuation and accounting method based on AI large model technology; Scenario: Disaster recovery system deployment at a fund company Media deployment and execution process (such as) Figure 5 (As shown).

[0104] S1. Disaster recovery environment startup (after system crash) 1. Media Insertion: Connect the encrypted SSD pre-installed with the ChatFA system to the backup server; 2. Automatic execution: Read the program code for media partition 1 and load it into memory; Start log recovery: Read the most recent operation sequence number from partition 2, for example: CT20250806; Rebuild the task chain: Restore the "Science and Technology Innovation Bond Fund Tiered" task to the interruption point.

[0105] S2. Execute intelligent valuation and accounting methods based on AI large-scale model technology. 1. Intent recognition: Parse cached instructions: "Continue processing Science and Technology Innovation Bond Fund Tiered Management" → Match product tiering agent; 2. Document parsing: Read rules from partition 3 → Extract the "Grading Instructions" field: "Priority ratio: 65% (user-corrected value before disaster recovery)"; 3. API call: Token authentication: Generates a temporary token using the key stored in the medium; Call the valuation system interface: / create-tier?tier=A; 4. Report generation: Output disaster recovery execution report (stored to partition 2): "[Disaster Recovery] Successfully created a priority A-level account set." Exception log: None (rulebase version v2.5).

[0106] S3. Media Security Mechanism 1. Storage encryption: AES-256 encrypted partition 2 log (compliant with the Financial Industry Information Security Specification JR / T 0071-2020); 2. Access Control: Two-factor authentication: physical media key + administrator biometrics; 3. Self-destruct mechanism: Three consecutive authentication failures trigger data overwrite.

[0107] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0108] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An intelligent valuation and accounting method based on AI large-scale model technology, characterized in that, include: It receives natural language commands input by the user and determines the target Agent through a dual-path matching mechanism of similarity retrieval from the historical dialogue database and inference from a large model. The financial business process rule base is invoked to break down user instructions into a chain of sub-tasks, including document upload, information extraction, API call, and data verification. Parse user-uploaded financial documents and use large model entity recognition and relationship extraction technologies to extract key fields of basic product information, fee structure, and account information; The extracted key fields are mapped to the valuation system interface parameters, a call request with a unique operation serial number is generated, and the operation log is recorded. Execution of anomaly classification: When a core field conflict or missing value is detected, the task is interrupted and manual intervention is requested; non-core field anomalies are handled later. Generate a visual summary report, outputting task execution status, exception handling trajectory, and a list of pending items.

2. The intelligent valuation and accounting method based on AI large model technology according to claim 1, characterized in that, The dual-path matching mechanism includes: The Agents whose similarity to the current user's command exceeds a preset threshold are retrieved from the historical dialogue database and selected as the first candidates. All Agent description information is input into the large model for inference matching as the second candidate; Prioritize selecting agents with consistent matching results from two different paths; otherwise, select candidate agents with higher similarity.

3. The intelligent valuation and accounting method based on AI large model technology according to claim 1, characterized in that, The information extraction steps include: When key fields are missing, data is completed across multiple uploaded documents according to preset priority; It supports users to define custom field extraction rules and dynamically add new rules to the financial business process rule library.

4. The intelligent valuation and accounting method based on AI large model technology according to claim 1, characterized in that, The specific anomaly classification processing is as follows: Level 1 anomaly: Product name or account information conflict or missing, forcibly interrupt the task and pop up a manual confirmation window; Level 2 anomaly: Rate format error or value exceeds limit, allowing users to choose immediate processing or delayed processing; Level 3 Exception: Non-critical field warning, logs exception information but does not interrupt the task flow.

5. The intelligent valuation and accounting method based on AI large model technology according to claim 1, characterized in that, The interface call steps include: Perform rule and logic validation on the extracted fields to ensure that the data conforms to the valuation system interface specifications; A user token authentication mechanism is used to verify the identity of the call request. Automatic retry when API call times out; the number of retries is configurable.

6. The intelligent valuation and accounting method based on AI large model technology according to claim 1, characterized in that, The visual summary report is a structured Excel file, containing: The completed operation list specifies the operation name, the fields involved, and the execution result. Exception handling records, noting conflicting fields, handling methods, and operation times; The pending items are categorized and displayed in groups based on missing data and rule conflicts.

7. An intelligent valuation and accounting system based on AI large-scale model technology, characterized in that: include: S1. Intent recognition module, used to parse natural language commands and match the target Agent; S2. Multi-Agent Collaboration Engine: Performs task decomposition, document parsing, and API calls; S3. Dynamic rule base, storing financial business process rules and custom extraction rules; S4. Audit feedback module, which generates operation logs and visual summary reports.

8. The intelligent valuation and accounting system based on AI large model technology according to claim 7, characterized in that, The multi-agent collaboration engine includes: The task decomposer breaks down sub-task chains based on a financial business process rule base. The document parser utilizes the entity recognition and relationship extraction capabilities of large model systems to extract key fields. The exception handler implements a three-level exception classification strategy and initiates a human-machine collaboration window.

9. The intelligent valuation and accounting system based on AI large model technology according to claim 7, characterized in that, The dynamic rule base supports: When a new financial document type is added, the field extraction logic is extended; When regulatory policies change, business rules are automatically updated.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the intelligent valuation and accounting method based on AI large model technology as described in any one of claims 1 to 6.

Images