An intelligent accounting interaction system and method based on a multi-level decision model and a generative UI

Abstract

The application belongs to the technical field of financial accounting and accounting, and relates to an intelligent accounting interaction system and method based on a multi-level decision model and a generative UI. The system comprises: an intelligent input arrangement module, which is used for receiving multi-modal input and standardizing unstructured data into structured objects; an intention routing module, which judges the user's task purpose according to the structured objects, and distributes the task to a voucher generation expert Agent, an optimized voucher expert Agent or a consultation problem expert Agent; wherein the voucher generation expert Agent adopts a pyramid decision model to execute voucher generation reasoning logic and outputs a decision result; a generative UI rendering engine, which configures an accounting atomic component library, receives the Agent decision result, dynamically calls the corresponding UI code block, and dynamically renders an interactive operation interface in the dialogue flow. Preferably, a continuous integration module continuously judges whether there is the same test when a test task is triggered, and if there is, the test is stopped, and if not, the test is performed.

Description

Technical Field

[0001] The present invention relates to the technical field of financial accounting, and belongs to an intelligent accounting interaction system and method based on a multi-level decision model and a generative UI. Background Art

[0002] In the field of corporate financial accounting, the generation of accounting vouchers is a core business process. Financial staff need to determine the accounting accounts, debit and credit directions and amounts based on the business content of original documents (such as OA approval forms, invoices, bank receipts, etc.), and finally enter the information into the financial system. With the development of SaaS technology, many financial software have begun to try to automate this process.

[0003] The prior art usually adopts a matching method "based on rule templates". That is, users preset a fixed rule that "document type A corresponds to debit account X plus credit account Y". When the system receives a document, it triggers the rule through keyword matching to generate a voucher.

[0004] However, the prior art has the following technical defects: 1. Rigid rules and lack of flexibility: Once the business scenario is slightly adjusted (for example, a special reimbursement situation for the same type of expense), the fixed rules cannot be adapted, and manual modification or reconfiguration is required.

[0005] 2. Inability to process natural language intentions: Users cannot temporarily intervene in the generation logic through colloquial instructions (such as "This meal expense is a benefit for employees and should be recorded as welfare expense"), and the system can only rigidly process it according to "meal invoice = business entertainment expense".

[0006] 3. Lack of adaptive learning ability: After the user manually modifies the generated result, the system cannot automatically learn the modification logic, which leads to the need for repeated modification when encountering the same situation next time.

[0007] 4. Difficult cold start: New users need to configure a large number of rules before use, which has an extremely high threshold.

[0008] 5. Separation of interaction and business: Traditional general-purpose AI assistants (such as Deerseek) can only return text or static Markdown tables when processing accounting services. When facing AI suggestions, financial staff still need to manually enter data in the ERP menu, which means AI only stays at the level of "consulting advisor" and cannot become a "business helper" that can operate directly, and the operation link is not closed. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides an intelligent accounting interaction system and method based on a multi-level decision model and a generative UI. It aims to solve the problems of existing financial software in voucher generation, such as cumbersome rule configuration, inability to understand users' temporary natural language commands, and lack of self-evolutionary capabilities based on historical data. By constructing an AI system that includes intent routing, a multi-level decision pyramid (L0-L3), and a feedback loop, it achieves high-precision, interventionist, and self-evolving automatic voucher generation.

[0010] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides an intelligent accounting interaction system based on a multi-level decision-making model and a generative UI, comprising: Intelligent input orchestration module: used to receive multimodal input, standardize unstructured data into structured objects, and extract user real-time instructions, business object metadata and user status information from the multimodal input; The user status recognition and permission verification module is used to determine whether a user has the formal account set operation permission based on the user account, account set binding relationship and account set access permission, and output the formal account set processing instruction or the temporary account set sandbox processing instruction. Intent routing module: Based on the structured object and the user's real-time instructions, it determines the user's task purpose and distributes the task to the corresponding Agent for processing. The Agent includes a voucher generation expert Agent, a voucher optimization expert Agent, and a consultation expert Agent. Among them, the voucher generation expert Agent is used to process voucher generation intents, the voucher optimization expert Agent is used to process intents for checking, optimizing, and correcting generated vouchers, and the consultation expert agent is used to process inquiry intents related to accounting, taxation, or system operation. Voucher Generation Expert Agent: Employs a pyramid decision model to execute the reasoning logic for voucher generation, generating a set of accounting data fields including summary, accounting subject, debit / credit direction, amount, auxiliary accounting, tax amount, voucher attachments, and business object references; Optimization Voucher Expert Agent: This agent receives generated vouchers or voucher drafts and, based on accounting rules, tax rules, historical voucher references, and system verification results, performs field consistency checks, debit / credit balance checks, account compatibility checks, and generates correction suggestions for the generated vouchers or voucher drafts. Consultation Expert Agent: Used to receive accounting, tax or system operation questions from users, call at least one information source from the legal knowledge base, accounting knowledge base, online search tools or internal enterprise knowledge base to generate consultation answers, and return the consultation answers to the dialogue flow; Generative UI rendering engine: Configured with a pre-installed accounting atomic component library in the enterprise accounting system, it receives structured results output by each Agent, dynamically calls corresponding UI code blocks from the accounting atomic component library according to the task type and structured results, maps the set of accounting data fields to component parameters, and directly renders interactive voucher cards, voucher optimization result cards, or consultation response cards in the dialogue flow; wherein, the interactive voucher card is the specific form of the interactive operation interface, used to display accounting data fields, receive user corrections, and return field differences.

[0011] Preferably, when the user status recognition and permission verification module outputs a temporary account sandbox processing instruction, the system introduces a temporary account sandbox mechanism to build a temporary account and virtual accounting entity for users without permissions or without an account in an isolated environment, allowing users to experience the voucher generation logic in the temporary account. After the user creates a formal account or obtains formal account permissions, the rule data, habit data and business object data generated in the temporary account are migrated to the formal account.

[0012] Preferably, the pyramid decision model follows a four-level decision hierarchy priority logic when generating each voucher field, including summary, accounting subject, debit / credit direction, amount, auxiliary accounting, tax amount, voucher attachments, and business object references. Specifically, this includes: L0 Instant Instruction: Set as the highest adjudication level, it directly adopts the mandatory instructions entered by the user, overriding other rules and historical references, and is used to determine at least one field among summary, accounting subject, auxiliary accounting, accounting period, or business attribution; L1 explicit rules: These rules are derived from pre-set industry rules, enterprise account configurations, charts of accounts, summary templates, business mapping tables, manually maintained rules, and confirmed rules verified and accumulated by Insight Loop. Specific weights are used in decision-making, and multi-condition combination rules have higher priority than general rules. L2 Historical Reference: When L0 and L1 cannot cover the data, the system uses enhanced retrieval generation technology to find similar historical vouchers, mimicking their summaries, accounting subjects, auxiliary accounting and tax processing habits to fill the rule gaps; L3 system fallback: When no valid information is available at the above levels, compliant voucher fields are generated by default based on the system's built-in industry standards; Decisions are made layer by layer in the order of L0, L1, L2, and L3.

[0013] Preferably, the system employs the Insight Loop user correction feedback learning mechanism, which records the field differences before and after the interactive voucher card is manually modified by the user. The field differences are attributed to at least one target field among the summary, accounting subject, auxiliary accounting, tax amount, accounting period, or business attribution. After legality verification and user confirmation, the rule base in the L1 explicit rules is expanded so that subsequent identical or similar tasks will preferentially hit the rule.

[0014] Preferably, the L1 explicit rules include L1-Structure configuration and L1-Content business rules; the L1-Structure configuration includes at least a summary template, date merging strategy, voucher entry structure and auxiliary accounting dimension configuration; the L1-Content business rules include at least a mapping rule between business type and accounting subject, tax processing rule, counterparty rule, expense attribution rule and manual maintenance rule.

[0015] Preferably, the L2 historical reference uses retrieval enhancement generation technology to retrieve historical voucher data; specifically, it extracts the task type, business type, amount, counterparty, tax attribute, time feature and account set identifier of the current business object, generates a query vector, and filters it according to account set, task type, business type and time range. After recalling the top-N historical vouchers through vector similarity, it is re-sorted, and the top-3 historical vouchers are selected, and their summary, accounting subject, auxiliary accounting and tax amount fields are extracted as L2 historical references.

[0016] Preferably, the accounting atomic component library in the generative UI rendering engine includes at least one of the following: voucher preview card, tax risk warning component, habit confirmation pop-up, voucher optimization result card, and consultation reply card; the generative UI rendering engine determines the component identifier based on the task type output by the intent routing module, determines the component parameters based on the structured results output by the Agent, and sends the user's confirmation, correction, or accounting operations in the component back to the corresponding Agent or feedback learning mechanism. On the other hand, the present invention also provides an interaction method applicable to the above-mentioned intelligent accounting interaction system based on a multi-level decision model and a generative UI, comprising the following steps: S1. The system receives the original object and user input, identifies the user status and account permissions, standardizes the original object and user input into a structured object, and extracts the task type, business object metadata and user real-time instructions. S2. Based on the task type, perform intent routing and distribute the credential generation task, credential optimization task, or consultation question task to the credential generation expert agent, credential optimization expert agent, or consultation question expert agent, respectively. S3. When the task type is a voucher generation task, perform a multi-threaded parallel full search on the structured object. The search includes at least L1 rule search, L2 historical voucher search and L3 industry base table search. S4. Based on the search results obtained in step S3 and the user's immediate instructions, a set of accounting data fields is generated by performing layer-by-layer coverage reasoning through the pyramid decision model. S5. Based on the current task type, retrieve the corresponding visualization module from the accounting atomic component library, inject the accounting data field set or the structured result output by the corresponding Agent into the component parameters, and render interactive voucher cards, voucher optimization result cards or consultation reply cards in real time on the client. S6. Receive confirmation, correction or posting operations from the user in the interactive voucher card, record field differences when field correction occurs and trigger the Insight Loop user correction feedback learning mechanism; when the user does not have formal account set permissions or no account set, limit the processing to the temporary account set sandbox, and migrate the temporary account set data to the formal account set after the migration conditions are met.

[0017] Preferably, step S3, the multi-threaded parallel full search, specifically includes the following parallel branches: Obtain the L1-Structure skeleton configuration; Retrieve the business rules of the hit L1-Content sorted by weight; A query vector is generated based on the characteristics of the current business object. After filtering by account set, task type, business type and time range, the Top-N historical vouchers are retrieved and reordered to obtain the Top-3 L2 historical vouchers. Obtain L3 industry baseline data.

[0018] Compared with existing technologies, this invention provides an intelligent accounting interaction system and method based on a multi-level decision-making model and a generative UI, which has the following beneficial effects: 1. This invention, through the L0-L3 model, not only ensures the seriousness of accounting standards (L1 / L3) but also empowers users with the ability to flexibly intervene through natural language (L0), thus solving the problem of rigidity in traditional rule engines.

[0019] 2. Through the InsightLoop mechanism, the system can learn from every manual correction made by the user and automatically expand the L1 rule base. As usage time goes on, the automation rate gradually approaches 100%.

[0020] 3. With the integration of RAG technology (L2), new users do not need to pre-configure a large number of rules. The system can directly refer to the industry practices of similar companies or a small amount of historical data from users to get started.

[0021] 4. This invention addresses users who haven't created a formal accounting system. The system dynamically constructs a "temporary accounting system" and a virtual accounting entity in memory. The rules (L1) and habitual data generated by the user in the temporary accounting system are seamlessly migrated to the formal database through an ETL process when the user decides to "create a formal accounting system." The temporary accounting system mechanism eliminates the psychological burden and data isolation costs for users trying out AI functions, ensuring a continuous experience.

[0022] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of an intelligent accounting interaction system based on a multi-level decision-making model and a generative UI, according to the present invention. Figure 2 This is a flowchart of the reasoning process within the credential generation expert agent of this invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0025] See Figures 1-2 A smart accounting interaction system based on a multi-level decision-making model and a generative UI, the testing device comprising: Intelligent input orchestration module: used to receive multimodal input (OA approval forms, transaction logs, invoices, natural language commands) and standardize unstructured data into structured objects.

[0026] Intent routing module: Based on the user's input of natural language or operation behavior, determine the user's intent (such as generating credentials, optimizing credentials, or asking questions), and distribute the task to the corresponding Agent.

[0027] Voucher Generation Expert Agent (Core): Responsible for executing the reasoning logic for voucher generation, using the L0-L3 pyramid decision model; Optimization Voucher Expert Agent: This agent receives generated vouchers or voucher drafts and, based on accounting rules, tax rules, historical voucher references, and system verification results, performs field consistency checks, debit / credit balance checks, account compatibility checks, and generates correction suggestions for the generated vouchers or voucher drafts. Consultation Expert Agent: Used to receive accounting, tax or system operation questions from users, call at least one information source from the legal knowledge base, accounting knowledge base, online search tools or internal enterprise knowledge base to generate consultation answers, and return the consultation answers to the dialogue flow; Feedback and Evolution Module: Captures user modification behavior and transforms it into new explicit rules.

[0028] Temporary account sandbox mechanism: Allows users without permissions or accounts to experience the generation logic in an isolated environment and supports seamless data migration.

[0029] Generative UI rendering engine: This is the core of the front-end interaction. The system comes pre-built with a highly cohesive "accounting atomic component library" (such as: voucher preview card, tax risk warning bar, habit confirmation pop-up).

[0030] This engine is responsible for receiving the decision results from the Agent. Instead of outputting a simple text stream, it dynamically calls the corresponding UI code blocks from the component library and renders an interactive interface directly in the dialogue flow.

[0031] In one possible implementation, the system strictly follows the following priority decision logic (from high to low) when generating each field (summary, account, auxiliary accounting) of the voucher: L0: Real-time Commands – "God's-eye View" Definition: Natural language instructions entered by the user at the moment of generating the operation (e.g., "This must be recorded in October").

[0032] Logic: It has the highest decision-making authority. Regardless of any pre-set rules or historical habits within the system, L0 must be satisfied first. Implementation: Extract L0_Instruction from the Prompt and force the subsequent logic to be overridden.

[0033] L1: Explicit Rules – “Mandatory Contracts” Definition: Hard constraints that are explicitly configured in the system.

[0034] L1-Structure (Skeleton Configuration): Global summary templates (such as [Date] + [Merchant]), date merging strategies.

[0035] L1-Content (Business Rules): Strong logic judgment for specific business content (such as "Didi Chuxing" automatically matching "travel expenses").

[0036] Conflict strategy: Employ specific weights, giving higher priority to rules with more specific conditions (combinations of multiple conditions) than general rules.

[0037] L2: Historical Reference – "Case Study" Definition: Historical credential data retrieved based on RAG (Retrieval Enhancement Generation) technology.

[0038] Logic: When L0 and L1 cannot cover the system (e.g., a new supplier without configured rules), the system uses vector retrieval to find the most similar Top-3 historical vouchers, mimicking their subject and summary habits.

[0039] Function: To identify and fill gaps, and to simulate user "habits".

[0040] L3: System safety net – "Public order and good morals" Definition: The system's built-in industry-standard default mapping table.

[0041] Logic: When there is no valid information at all the above levels, a standardized default generation is performed according to accounting standards.

[0042] In one possible implementation, the workflow of the above system is as follows: 1. Standardized tagging: The system receives the original object (such as OA document), identifies its type, and extracts metadata.

[0043] 2. Parallel Full Search: For this task, the following threads are executed in parallel: Obtain the L1-Structure skeleton configuration.

[0044] Retrieve the list of hit L1-Content business rules (sorted by weight).

[0045] Vector retrieval of Top-3 L2 historical credentials.

[0046] Obtain L3 industry baseline data.

[0047] 3. Pyramid Reasoning: The search results are injected into the structured Prompt, forcing the LLM (Large Language Model) to execute "layer-by-layer coverage" logic. If an L0 instruction exists, it is explicitly required in the Prompt to cover the L1 / L2 results.

[0048] 4. Intent-driven atomic component assembly: Component retrieval: AI retrieves the corresponding visualization module (such as Voucher_Preview_Card) from the component library based on the current task type (such as "Generate Voucher").

[0049] Data population: Inject the accounting data (summary, account, amount) inferred from the L0-L3 model into the component parameters.

[0050] Streaming rendering: The dialog box instantly renders a WYSIWYG voucher card. Users can complete the business loop simply by clicking the "One-Click Transaction" button on the card without leaving the dialog window.

[0051] 5. Temporary Account Sets and Migration Mechanism For users who have not created a formal accounting system, the system dynamically constructs a "temporary accounting system" and a virtual accounting entity in memory.

[0052] The rules (L1) and habit data generated by users in the temporary account set are seamlessly migrated to the official database through the ETL process when the user decides to "create the official account set", ensuring the continuity of the experience.

[0053] In one possible implementation, decision-making hierarchy can be replaced: the number of levels can be reduced (e.g., only rules and history are retained) or the number of levels can be increased (e.g., departmental rules are added), but the core logic of "instructions covering rules and rules covering history" remains unchanged.

[0054] In one possible implementation, model substitution is possible: the core inference model can use Qwen, Deepseek, or a finely tuned private large model. In one possible implementation, vector retrieval can be used as an alternative: L2-level historical retrieval can use traditional keyword full-text retrieval instead of vector retrieval, but the effect will be reduced in terms of semantic similarity matching.

[0055] In one possible implementation, the "taxi fare" scenario is handled as follows: 1. Input: The user uploads a "Didi Chuxing" invoice and enters the L0 command: "Make a voucher for this taxi invoice taken after working overtime".

[0056] 2. Search: (1) The L1 rule base contains "Didi Chuxing > Travel Expenses".

[0057] (2) In the history of L2, there is “Didi Chuxing > transportation fee”.

[0058] 3. Reasoning: (1) L1 hit “travel expenses”.

[0059] (2) But in the L0 stage, the system recognizes the intention to "work overtime".

[0060] (3) Combine L3 or common sense to link “overtime taxi” to “management expenses - welfare expenses” or specific items.

[0061] (4) Interactive performance: Instead of replying "OK, generated for you..." in the dialog box, the system directly calls the Voucher_Card component to render a visual voucher card at the bottom of the current dialog flow.

[0062] (5) Card content: Display summary "overtime taxi fare", account "management expenses - welfare expenses", amount "2000 yuan", and highlight the debit and credit balance status.

[0063] (6) Operation closed loop: After the user browses the card and finds it correct, he / she can directly click the "One-click entry" button in the lower right corner of the card. The system background calls the API to complete the formal voucher writing and updates the card status to "entered" on the interface.

[0064] 4. Feedback: If the user confirms that the rule is correct, the system does not need to add a new rule. If the user changes the rule to "Project Cost", the system will prompt the user to save the new rule "Overtime Taxi = Project Cost".

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent accounting interaction system based on a multi-level decision-making model and a generative UI, characterized in that: include: Intelligent input orchestration module: used to receive multimodal input, standardize unstructured data into structured objects, and extract user real-time instructions, business object metadata and user status information from the multimodal input; The user status recognition and permission verification module is used to determine whether a user has the formal account set operation permission based on the user account, account set binding relationship and account set access permission, and output the formal account set processing instruction or the temporary account set sandbox processing instruction. Intent routing module: Based on the structured object and the user's real-time instructions, it determines the user's task purpose and distributes the task to the corresponding Agent for processing. The Agent includes a voucher generation expert Agent, a voucher optimization expert Agent, and a consultation expert Agent. Among them, the voucher generation expert Agent is used to process voucher generation intents, the voucher optimization expert Agent is used to process intents for checking, optimizing, and correcting generated vouchers, and the consultation expert agent is used to process inquiry intents related to accounting, taxation, or system operation. Voucher Generation Expert Agent: Employs a pyramid decision model to execute the reasoning logic for voucher generation, generating a set of accounting data fields including summary, accounting subject, debit / credit direction, amount, auxiliary accounting, tax amount, voucher attachments, and business object references; Optimization Voucher Expert Agent: This agent receives generated vouchers or voucher drafts and, based on accounting rules, tax rules, historical voucher references, and system verification results, performs field consistency checks, debit / credit balance checks, account compatibility checks, and generates correction suggestions for the generated vouchers or voucher drafts. Consultation Expert Agent: Used to receive accounting, tax or system operation questions from users, call at least one information source from the legal knowledge base, accounting knowledge base, online search tools or internal enterprise knowledge base to generate consultation answers, and return the consultation answers to the dialogue flow; Generative UI rendering engine: Configured with a pre-installed accounting atomic component library in the enterprise accounting system, it receives structured results output by each Agent, dynamically calls corresponding UI code blocks from the accounting atomic component library according to the task type and structured results, maps the set of accounting data fields to component parameters, and directly renders interactive voucher cards, voucher optimization result cards, or consultation response cards in the dialogue flow; wherein, the interactive voucher card is the specific form of the interactive operation interface, used to display accounting data fields, receive user corrections, and return field differences.

2. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 1, characterized in that: When the user status recognition and permission verification module outputs a temporary account sandbox processing instruction, the system introduces a temporary account sandbox mechanism to build a temporary account and virtual accounting entity for users without permissions or without an account in an isolated environment. This allows users to experience the voucher generation logic in the temporary account. After the user creates a formal account or obtains formal account permissions, the rule data, habit data, and business object data generated in the temporary account are migrated to the formal account.

3. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 1, characterized in that: The pyramid decision-making model follows a four-level priority logic when generating each voucher field, including summary, accounting subject, debit / credit direction, amount, auxiliary accounting, tax amount, voucher attachments, and business object references. Specifically, it includes: L0 Instant Instruction: Set as the highest adjudication level, it directly adopts the mandatory instructions entered by the user, overriding other rules and historical references, and is used to determine at least one field among summary, accounting subject, auxiliary accounting, accounting period, or business attribution; L1 explicit rules: These rules are derived from pre-set industry rules, enterprise account configurations, charts of accounts, summary templates, business mapping tables, manually maintained rules, and confirmed rules verified and accumulated by Insight Loop. Specific weights are used in decision-making, and multi-condition combination rules have higher priority than general rules. L2 Historical Reference: When L0 and L1 cannot cover the data, the system uses enhanced retrieval generation technology to find similar historical vouchers, mimicking their summaries, accounting subjects, auxiliary accounting and tax processing habits to fill the rule gaps; L3 system fallback: When no valid information is available at the above levels, compliant voucher fields are generated by default based on the system's built-in industry standards; Decisions are made layer by layer in the order of L0, L1, L2, and L3.

4. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 3, characterized in that: The system employs the Insight Loop user correction feedback learning mechanism, which records the field differences before and after the interactive voucher card is manually modified by the user. The field differences are attributed to at least one target field among the summary, accounting subject, auxiliary accounting, tax amount, accounting period, or business attribution. After legality verification and user confirmation, the rule base in the L1 explicit rules is expanded so that subsequent identical or similar tasks will preferentially hit the rule.

5. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 3, characterized in that: The L1 explicit rules include L1-Structure configuration and L1-Content business rules; the L1-Structure configuration includes at least a summary template, date merging strategy, voucher entry structure and auxiliary accounting dimension configuration; the L1-Content business rules include at least a mapping rule between business type and accounting subject, tax processing rule, counterparty rule, expense attribution rule and manual maintenance rule.

6. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 3, characterized in that: In the L2 historical reference, historical voucher data is retrieved using retrieval enhancement generation technology. Specifically, the task type, business type, amount, counterparty, tax attribute, time feature, and account set identifier of the current business object are extracted to generate a query vector. The vector is then filtered by account set, task type, business type, and time range. After recalling the Top-N historical vouchers through vector similarity, the data is re-ranked, and the Top-3 historical vouchers are selected. Their summary, accounting subject, auxiliary accounting, and tax amount fields are extracted as L2 historical references.

7. The intelligent accounting interaction system based on a multi-level decision-making model and a generative UI according to claim 1, characterized in that: The accounting atomic component library in the generative UI rendering engine includes at least one of the following: voucher preview card, tax risk warning component, habit confirmation pop-up, voucher optimization result card, and consultation reply card; the generative UI rendering engine determines the component identifier based on the task type output by the intent routing module, determines the component parameters based on the structured results output by the Agent, and sends the user's confirmation, correction, or accounting operations in the component back to the corresponding Agent or feedback learning mechanism.

8. The interaction method of an intelligent accounting interaction system based on a multi-level decision model and a generative UI according to any one of claims 1-7, characterized in that: Includes the following steps: S1. The system receives the original object and user input, identifies the user status and account permissions, standardizes the original object and user input into a structured object, and extracts the task type, business object metadata and user real-time instructions. S2. Based on the task type, perform intent routing and distribute the credential generation task, credential optimization task, or consultation question task to the credential generation expert agent, credential optimization expert agent, or consultation question expert agent, respectively. S3. When the task type is a voucher generation task, perform a multi-threaded parallel full search on the structured object. The search includes at least L1 rule search, L2 historical voucher search and L3 industry base table search. S4. Based on the search results obtained in step S3 and the user's immediate instructions, a set of accounting data fields is generated by performing layer-by-layer coverage reasoning through the pyramid decision model. S5. Based on the current task type, retrieve the corresponding visualization module from the accounting atomic component library, inject the accounting data field set or the structured result output by the corresponding Agent into the component parameters, and render interactive voucher cards, voucher optimization result cards or consultation reply cards in real time on the client. S6. Receive confirmation, correction or posting operations from the user in the interactive voucher card, record field differences when field correction occurs and trigger the Insight Loop user correction feedback learning mechanism; when the user does not have formal account set permissions or no account set, limit the processing to the temporary account set sandbox, and migrate the temporary account set data to the formal account set after the migration conditions are met.

9. The intelligent accounting interaction method based on a multi-level decision-making model and a generative UI according to claim 8, characterized in that: In step S3, the multi-threaded parallel full search specifically includes the following parallel branches: Obtain the L1-Structure skeleton configuration; Retrieve the business rules of the hit L1-Content sorted by weight; A query vector is generated based on the characteristics of the current business object. After filtering by account set, task type, business type and time range, the Top-N historical vouchers are retrieved and reordered to obtain the Top-3 L2 historical vouchers. Obtain L3 industry baseline data.

Images