Layered intelligent reasoning system with double auditing interaction

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a layered intelligent reasoning system with double auditing interaction, which comprises a high-level planning module and a task execution subsystem. By adopting a nested architecture design of a main workflow and a sub-workflow, integrating a dual manual collaboration mechanism and a dynamic process control capability, utilizing a large language model as a core inference engine, and integrating with an external tool service through an MCP protocol, full-process controllable execution from task understanding to data acquisition, analysis and reporting is realized; in a professional scene with a high reliability requirement, a user is allowed to directly intervene in disassembling logic in a task refinement stage, invalid data calling and reasoning deviation caused by misunderstanding of artificial intelligence are avoided, overall examination and approval of a first-level man-machine interaction interface and local correction of a second-level man-machine interaction interface are combined, and the reliability of the user is improved. A complete man-machine co-control chain from macroscopic to microscopic is formed, and the accuracy and credibility of the final conclusion of task execution are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a hierarchical intelligent reasoning system with double auditing interaction. BACKGROUND

[0002] Artificial intelligence is a branch of computer science, which is an intelligent system that reacts in a similar way to human intelligence through intelligent reasoning models. At present, it is mainly applied in the form of AI assistant question and answer software, which can answer user questions and help users create texts and generate pictures. In simple terms, an AI assistant is an intelligent reasoning system that can understand problems, logically deduce, and give reasoned conclusions. However, in professional scenarios with high reliability requirements such as financial investment and research, insurance services, and government consultation, users often need to complete comprehensive tasks involving multi-step analysis, cross-system data calling, and complex logical reasoning. Current mainstream AI intelligent reasoning systems mostly use linear or tree-like thinking chain methods for reasoning, which can complete basic task decomposition but expose many problems in actual application, resulting in low accuracy and reliability of the final conclusion.

[0003] The current intelligent reasoning system has a rigid and uncontrollable reasoning process. Once started, it runs as a black box and cannot be interrupted, jumped or rolled back. Errors are difficult to correct after accumulation, and there is a lack of pre-audit mechanism. AI automatically decomposes high-level tasks into specific operation requests, which may deviate from the user's true intention if the understanding is biased. Most intelligent reasoning systems only allow corrections after the results are generated, which is a post-repair method and cannot prevent the occurrence of invalid or redundant operations. Moreover, the task decomposition results generated by AI usually go directly to the execution stage, and users cannot adjust or optimize them. Especially when facing semantic ambiguity or multi-solution problems, AI is likely to make a decomposition solution that does not meet business needs, which seriously affects the accuracy and reliability of the final conclusion. SUMMARY

[0004] Therefore, the present application provides a hierarchical intelligent reasoning system with double auditing interaction to solve the problem that the current intelligent reasoning technology in professional scenarios with high reliability requirements is prone to deviate from user business needs in task decomposition, resulting in low accuracy and reliability of the final conclusion.

[0005] To achieve the above purpose, the present application provides a hierarchical intelligent reasoning system with double auditing interaction, which comprises a high-level planning module and a task execution subsystem. The high-level planning module is used to receive user task requests sent by the user end and call a large language model to analyze user intentions and generate stage execution strategies based on a pre-defined structured planning prompt template. The high-level planning module is connected to the first-level human-computer interaction interface. The high-level planning module can determine whether the information requested by the user task is sufficient based on the structured planning prompt template. When the information requested by the user task is insufficient, the high-level planning module will transfer the output to the first-level human-computer interaction interface to prompt the user to supplement information or confirm the execution direction. The task execution subsystem is used to break down the phased execution strategy of the high-level planning module into executable actions, generate a list of query tasks, coordinate data acquisition and analysis, and output execution results. The task execution subsystem is connected to a secondary human-computer interaction interface, which is used to display the list of query tasks generated by the task execution subsystem. Through human-computer interaction, the decomposition results are reviewed semantically, and specific query task items can be added, deleted, or modified.

[0006] Furthermore, it also includes a results summary module; The structured planning prompt template includes key information items and their completeness judgment criteria. The high-level planning module judges the contextual condition information items of the user task request based on the key information items and their completeness judgment criteria in the structured planning prompt template. If any of the contextual condition information items in the user task request is missing and cannot be derived, the high-level planning module determines that the information in the user task request is insufficient. The high-level planning module will output to the first-level human-computer interaction interface, waiting for the user to supplement information or confirm the execution direction. If the high-level planning module determines that the user's task request information is sufficient, it jumps to the result aggregation module. The result aggregation module calls the large language model to reason about the user's task request and integrates all intermediate results and reasoning clues from the high-level planning module and the task execution subsystem to generate a structured final report.

[0007] Furthermore, the structured planning prompt template can be customized and dynamically updated by users, who can add, modify, or delete key information items in the template according to specific business needs.

[0008] Furthermore, the task execution subsystem includes a task refinement module, which decomposes the phased execution strategy of the high-level planning module into specific tool call instructions based on a predefined execution refinement prompt template, and generates a query task list.

[0009] Furthermore, the task execution subsystem also includes an external service call module, which is connected to the task refinement module. The external service call module can call an external MCP client service to perform a query based on the query task list, and cache the returned data in a structured manner to generate query results.

[0010] Furthermore, the task execution subsystem also includes a data analysis module, which is connected to the external service call module. The data analysis module is used to clean, compare and reason about the query results and output the execution results.

[0011] Furthermore, the data analysis module is also connected to the task refinement module, and the data analysis module can determine whether all query tasks within the phased execution strategy have been completed. If all query tasks are completed, the data analysis module will output the execution results; If any query task is incomplete, the data analysis module will return the incomplete query task to the task refinement module for further breakdown, forming a local loop.

[0012] Furthermore, it also includes a status update module, which is connected to the task execution subsystem and the high-level planning module respectively. The status update module is used to receive the execution results output by the task execution subsystem and update the status of the structured planning prompt template in the high-level planning module according to the execution results.

[0013] Furthermore, the results aggregation module can provide report feedback, taking new requirements raised by users during the reporting phase as new input to re-trigger the high-level planning module and start the next round of reasoning process.

[0014] Furthermore, the high-level planning module, the task execution subsystem, and the result aggregation module are all equipped with a state coordination mechanism, which transmits context information through LangGraph's shared state mechanism and supports the persistence of intermediate states and non-sequential jumps across nodes.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a hierarchical intelligent reasoning system that integrates a dual human-computer collaboration mechanism and an editable thought chain structure. Through the refined control of the task execution process and the design of a semantic guidance mechanism, the reliability and practicality of the AI ​​system are improved. The hierarchical architecture design decouples high-level decision-making from low-level execution, facilitating independent module upgrades and cross-scenario reuse. It supports non-sequential process jumps based on state coordination mechanisms, breaking through the limitations of the traditional unidirectional irreversible reasoning chain and enhancing the system's fault tolerance. In professional scenarios with high reliability requirements, after the task is decomposed into multiple query steps, users can add or delete redundant items or modify the query scope in the secondary human-computer interaction interface to ensure that the execution path fully matches the actual needs. By allowing users to directly intervene in the decomposition logic during the task refinement stage, invalid data calls and reasoning deviations caused by AI misunderstandings are avoided. Combining the overall approval of the primary human-computer interaction interface and the local correction of the secondary human-computer interaction interface, a complete human-computer co-control chain from macro to micro is formed, significantly improving the accuracy and credibility of the final conclusion of task execution. Attached Figure Description

[0016] Figure 1 This is a system block diagram of the hierarchical intelligent reasoning system with dual-audit interaction in this embodiment; Figure 2 This is an application flowchart of the hierarchical intelligent reasoning system with dual-audit interaction in this embodiment. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0020] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" 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.

[0021] Please see Figure 1 As shown, this embodiment provides a hierarchical intelligent reasoning system with dual review interaction. It adopts a nested architecture design of "main workflow + sub-workflow", integrates dual human collaboration mechanism and dynamic process control capability. The system builds an editable state graph workflow based on LangGraph, uses the Large Language Model (LLM) as the core reasoning engine, and integrates with external tool services through the MCP (Model Control Protocol) to achieve fully controllable execution of the entire process from task understanding to data acquisition, analysis and reporting.

[0022] Specifically, the hierarchical intelligent reasoning system with dual-audit interaction in this embodiment includes: a high-level planning module, a task execution subsystem, a result aggregation module, and a status update module; The high-level planning module receives user task requests sent by the client, calls the large language model, parses the user intent based on the predefined structured planning prompt template, and generates phased execution strategies. The structured planning prompt template can define the required key information items and their completeness judgment criteria for different task types. The high-level planning module dynamically judges whether the contextual condition information items in the user's task request are sufficient based on the required key information items and their completeness judgment criteria. In this embodiment, different task types are mainly for professional scenarios with high reliability requirements, such as policy renewal assessment and asset allocation advice. Taking the policy renewal assessment task as an example, the structured planning prompt template sets the required key information items: the insured's age, health disclosure status, annual income, and risk preference. The high-level planning module judges the user's task request by calling the large language model. If any of the contextual condition information items in the user's task request are missing and cannot be derived, the high-level planning module will determine that the information in the user's task request is insufficient; otherwise, it will determine that the information in the user's task request is sufficient. The structured planning prompt template supports user customization and dynamic updates. Domain experts or system administrators can add, modify, or delete assessment rules in the template according to specific business needs such as the launch of new insurance products or the implementation of new regulatory rules, without changing the core system code. This mechanism enables the layered intelligent reasoning system with dual review interaction to flexibly adapt to various professional tasks and continuously evolve to support emerging scenarios.

[0023] The high-level planning module is connected to a first-level human-computer interaction interface. When the high-level planning module determines that the information in the user's task request is insufficient, the high-level planning module will pause the automated process and switch to the first-level human-computer interaction interface to wait for the user to supplement information or confirm the execution direction.

[0024] The first-level human-computer interaction interface can display the phased execution strategies generated by the high-level planning module, allowing users to review, modify, or approve the phased execution strategies. Users can adjust the task scope, supplement missing information, or directly reject unreasonable plans to ensure that subsequent execution conforms to the true business intent.

[0025] If the user still refuses to provide the key information items required by the structured planning prompt template during the human feedback stage of the first-level human-computer interaction interface, the high-level planning module will instruct the large model to generate a hypothetical report with a clear statement of uncertainty, and provide options to continue to complete the information or accept the current conclusion, ensuring that the process can proceed without losing rigor.

[0026] The task execution subsystem is a nested sub-workflow managed by LangGraph. It refines the phased execution strategies of the high-level planning module into executable actions and coordinates data acquisition and analysis. It includes... The task refinement module calls the large language model and, based on predefined execution refinement prompt templates, breaks down the phased execution strategy of the high-level planning module into specific tool call instructions, generating a list of query tasks; the specific tool call instructions are the MCP tool call instructions, including the tool name, parameter list and expected output format; The external service invocation module invokes external services to query results through the MCP client and caches the returned data in a structured manner; it supports HTTP streaming, STDIO, SSE and other methods. The data analysis module cleans, compares, and infers the query results; it also determines whether all sub-tasks within the phased execution strategy have been completed; if there are still incomplete items, it only returns to the task refinement module for re-decomposition and re-calling of the incomplete sub-tasks, forming a local loop to avoid repeatedly executing successfully completed tasks; if all sub-tasks are completed, the execution results are output.

[0027] The task refinement module is connected to the secondary human-computer interaction interface, which displays the list of query tasks generated by the task refinement module, allowing users to conduct semantic-level review of the breakdown results. Users can add, delete, or modify specific query items here, such as adding payment flexibility comparison and deleting competitor analysis.

[0028] The results summary module is used to call the large language model to reason about the user's task request when the high-level planning module determines that the information of the user's task request is sufficient, and to integrate all intermediate results and reasoning clues of the high-level planning module and the task execution subsystem to generate a structured final report. The results summary module also has a report feedback function. If a user raises a new request during the reporting stage, such as "Please add inflation adjustment analysis", the results summary module will take this feedback as a new input, re-trigger the high-level planning module, start a new round of reasoning process, and achieve multiple iterations.

[0029] The status update module is connected to the task execution subsystem and the high-level planning module respectively. The status update module is used to receive the execution results output by the task execution subsystem and update the status of the structured planning prompt template in the high-level planning module according to the execution results. That is, the structured planning prompt template is filled with execution result data to reduce the number of key information items required to determine whether the information is sufficient.

[0030] In this embodiment, the high-level planning module, task execution subsystem, and result aggregation module of the hierarchical intelligent reasoning system with dual-audit interaction are all equipped with a state coordination mechanism. Context information is passed through LangGraph's shared state mechanism, supporting the persistence of intermediate states. At the same time, the state coordination mechanism supports non-sequential jumps across nodes. For example, when execution fails, it jumps back to task refinement; when conditions are met, it terminates early; and when the user intervenes, it interrupts the process, thereby enabling real-time editing and flexible control of the reasoning chain.

[0031] Please continue reading. Figure 2 As shown, in this embodiment, the application process of the hierarchical intelligent reasoning system with dual-audit interaction is as follows: Users submit natural language requests, such as "Help me analyze whether this policy is worth renewing?"; The system requests access to the high-level planning module and determines whether the information is sufficient based on the structured planning prompt template; that is, it performs semantic understanding on the request and determines whether the existing information is sufficient to generate a conclusion. If all information is available, such as age, health status, policy ID, and income, then proceed to the results summary module for quick report generation. If insufficient, such as a lack of risk appetite, then proceed to the first-level human-computer interaction interface; After switching to the first-level human-computer interaction interface, the user is prompted to supplement the missing information or confirm the execution direction; If the user provides additional information, it is passed to the high-level planning module, which determines whether the information is sufficient based on the structured planning prompt template. If the user does not provide additional information but confirms the execution direction; The high-level planning module initiates the task execution subsystem; The task refinement module, based on predefined execution refinement prompt templates, breaks down the phased execution strategies of the high-level planning module into specific tool call instructions, generating a list of query tasks, such as "query the historical claims records of this policy" and "obtain the yield of similar products". The breakdown results are submitted to the secondary human-computer interaction interface, prompting the user to review and confirm whether the task breakdown is reasonable; the user can intervene at this stage, for example... A new query option has been added: "Also need to compare payment years"; Remove irrelevant items: "No need to check competitor information"; Revise the existing problem description to make it more accurate; After confirming that the secondary human-computer interaction interface is correct, the system continues to execute; The external service call module calls external services through the MCP client to query results based on the query task list, and caches the returned data in a structured manner. The data analysis module performs completion reasoning, comparison calculations, and conclusion summarization on the returned data; if all subtasks are completed, the execution results are output; if there are still incomplete items, only the incomplete subtasks are returned to the task refinement module for re-disassembly and re-call. The status update module receives the execution results output by the task execution subsystem and updates the status of the structured planning prompt template in the high-level planning module according to the execution results. The high-level planning module determines whether the information is sufficient based on the updated structured planning prompt template. If the information is sufficient, it jumps to the results summary module. The results summary module calls the large language model to reason about the user's task request, and integrates all intermediate results and reasoning clues from the high-level planning module and the task execution subsystem to generate a structured final report. This embodiment's hierarchical intelligent reasoning system with dual-review interaction forms a complete closed loop of "planning, dual review, step-by-step execution, iterative feedback, and report generation," possessing high controllability and adaptability. It divides the user task reasoning process into two levels: high-level planning and specific execution. A nested graph structure enables collaborative scheduling of the main workflow and sub-workflows, allowing complex tasks to achieve phased control and status feedback within a unified framework. A dual manual intervention mechanism is set up at key nodes in the task flow. The first level, located after the overall plan is generated, is used to review the feasibility of execution and supplement missing information. The second level, located after the specific operation is broken down and before tool invocation, is used to display the task decomposition results and allow users to add, delete, or modify them, achieving refined semantic correction of the reasoning path and forming a two-layer collaborative control system covering strategic decision-making and tactical execution. Build a process jump capability based on state coordination mechanism to support non-sequential migration across nodes at runtime according to control instructions, such as jumping back to planning after execution failure and directly connecting to report generation after information completion, thereby forming an editable, interruptible and recoverable intelligent reasoning chain.

[0032] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hierarchical intelligent reasoning system with dual-audit interaction, characterized in that, include: High-level planning module and task execution subsystem; The high-level planning module is used to receive user task requests sent by the user terminal, and call the large language model to parse the user intent based on the predefined structured planning prompt template and generate a phased execution strategy. The high-level planning module is connected to the first-level human-computer interaction interface. The high-level planning module can determine whether the information requested by the user task is sufficient based on the structured planning prompt template. When the information requested by the user task is insufficient, the high-level planning module will transfer the output to the first-level human-computer interaction interface to prompt the user to supplement information or confirm the execution direction. The task execution subsystem is used to break down the phased execution strategy of the high-level planning module into executable actions, generate a list of query tasks, coordinate data acquisition and analysis, and output execution results. The task execution subsystem is connected to a secondary human-computer interaction interface, which is used to display the list of query tasks generated by the task execution subsystem. Through human-computer interaction, the decomposition results are reviewed semantically, and specific query task items can be added, deleted, or modified.

2. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 1, characterized in that, It also includes a results summary module; The structured planning prompt template includes key information items and their completeness judgment criteria. The high-level planning module judges the contextual condition information items of the user task request based on the key information items and their completeness judgment criteria in the structured planning prompt template. If any of the contextual condition information items in the user task request is missing and cannot be derived, the high-level planning module determines that the information in the user task request is insufficient. The high-level planning module will output to the first-level human-computer interaction interface, waiting for the user to supplement information or confirm the execution direction. If the high-level planning module determines that the user's task request information is sufficient, it jumps to the result aggregation module. The result aggregation module calls the large language model to reason about the user's task request and integrates all intermediate results and reasoning clues from the high-level planning module and the task execution subsystem to generate a structured final report.

3. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 2, characterized in that, The structured planning prompt template can be customized and dynamically updated by users, who can add, modify, or delete key information items in the template according to specific business needs.

4. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 1, characterized in that, The task execution subsystem includes a task refinement module, which decomposes the phased execution strategy of the high-level planning module into specific tool call instructions based on a predefined execution refinement prompt template, and generates a query task list.

5. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 4, characterized in that, The task execution subsystem also includes an external service call module, which is connected to the task refinement module. The external service call module can call an external MCP client service to perform a query based on the query task list, and cache the returned data in a structured manner to generate query results.

6. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 5, characterized in that, The task execution subsystem also includes a data analysis module, which is connected to the external service call module. The data analysis module is used to clean, compare and reason about the query results and output the execution results.

7. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 6, characterized in that, The data analysis module is also connected to the task refinement module, and the data analysis module can determine whether all query tasks within the phased execution strategy have been completed. If all query tasks are completed, the data analysis module will output the execution results; If any query task is incomplete, the data analysis module will return the incomplete query task to the task refinement module for further breakdown, forming a local loop.

8. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 7, characterized in that, It also includes a status update module, which is connected to the task execution subsystem and the high-level planning module respectively. The status update module is used to receive the execution results output by the task execution subsystem and update the status of the structured planning prompt template in the high-level planning module according to the execution results.

9. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 2, characterized in that, The results aggregation module can provide report feedback, taking new requirements raised by users during the reporting phase as new input to re-trigger the high-level planning module and start the next round of reasoning process.

10. The hierarchical intelligent reasoning system with dual-audit interaction according to claim 9, characterized in that, The high-level planning module, the task execution subsystem, and the result aggregation module are all equipped with a state coordination mechanism. They use LangGraph's shared state mechanism to pass context information and support the persistence of intermediate states and non-sequential jumps across nodes.

Images