Business process processing method, apparatus, device, storage medium, and program product

Abstract

This application provides a business process processing method, apparatus, device, storage medium, and program product. The method includes: constructing a dynamic graph structure based on a preset business process, wherein the dynamic graph structure contains multiple nodes and edges defining the execution relationships between nodes, and the nodes include functional nodes, business nodes, and collaborative scheduling nodes; receiving a user request sent by a user terminal and initializing a global state based on the user request; executing an iterative scheduling process in the dynamic graph structure: determining the target node to be executed based on the current global state using a state-driven routing algorithm; executing the node logic corresponding to the target node to obtain the execution result; updating the global state based on the execution result; repeatedly executing the iterative scheduling process until a final response to the user request is generated; and sending the final response to the user terminal. This improves the flexibility, accuracy, and efficiency of business process processing.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a business process processing method, apparatus, device, storage medium, and program product. Background Technology

[0002] With the rapid development of the digital economy and the diversification and upgrading of customer service needs, intelligent customer service systems have become a core support for improving service efficiency and optimizing user experience.

[0003] Currently, intelligent customer service systems typically handle user requests through linear processes or predefined static rules.

[0004] However, existing technologies rely on linear processes or predefined static rules, resulting in low efficiency and insufficient accuracy in business processing. Summary of the Invention

[0005] This application provides business process processing methods, apparatus, equipment, storage media, and program products to improve business processing efficiency and accuracy.

[0006] Firstly, this application provides a business process processing method, including:

[0007] A dynamic graph structure is constructed based on a preset business process. The dynamic graph structure contains multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes.

[0008] Receive user requests sent by the client and initialize the global state according to the user requests;

[0009] In the dynamic graph structure, an iterative scheduling process is executed as follows: based on the current global state, a state-driven routing algorithm is used to determine the target node to be executed; the node logic corresponding to the target node is executed to obtain the execution result; the global state is updated according to the execution result; the iterative scheduling process is repeated until a final response to the user request is generated.

[0010] The final response is sent to the user terminal.

[0011] In one possible implementation, determining the target node to be executed based on the current global state using a state-driven routing algorithm includes: performing route matching based on preset business rules according to the current information in the global state to obtain a first matching result; if the first matching result is not empty, determining the first matching result as the target node to be executed; if the first matching result is empty, obtaining the current user request based on the current information in the global state; calculating the association score between each candidate node and the current user request; and determining the target node to be executed from the candidate nodes based on the association score.

[0012] In one possible implementation, calculating the association score between each candidate node and the current user request includes: calculating the semantic matching degree between each candidate node and the current user request; calculating the historical execution success rate of each candidate node; and performing weighted processing based on the semantic matching degree and the historical execution success rate to obtain the association score between each candidate node and the current user request.

[0013] In one possible implementation, determining the target node to be executed from the candidate nodes based on the association score includes: determining the candidate node with the highest association score as the target node to be executed; when multiple candidate nodes have the highest and the same association score, using a round-robin mechanism to determine the target node to be executed from the multiple candidate nodes.

[0014] In one possible implementation, the execution iterative scheduling process includes: determining a first cooperative scheduling node based on a first global state using a state-driven routing algorithm; the first global state being an initialized global state; performing deep analysis on the user request through the first cooperative scheduling node to obtain an intent classification label and a confidence level; performing feature extraction processing on the user request through the first cooperative scheduling node to obtain multi-dimensional feature data; calculating a complexity score through the first cooperative scheduling node based on the multi-dimensional feature data; and updating the first global state to a second global state through the first cooperative scheduling node based on the intent classification label, confidence level, and complexity score.

[0015] In one possible implementation, the execution of the iterative scheduling process includes: determining a second cooperative scheduling node based on the second global state using a state-driven routing algorithm; determining a decision result through the second cooperative scheduling node based on the complexity score and a preset routing decision function; and updating the second global state to a third global state based on the decision result.

[0016] In one possible implementation, determining the decision result based on the complexity score and a preset routing decision function includes: determining the relationship between the complexity score and a preset first complexity threshold and a second complexity threshold; wherein the first complexity threshold is less than the second complexity threshold; if the complexity score is less than the first complexity threshold, then determining the decision result for the single-threaded mode; if the complexity score is greater than the first complexity threshold and less than the second complexity threshold, then determining the decision result for the coroutine concurrent mode; if the complexity score is greater than the second complexity threshold, then determining the decision result for the distributed scheduling mode.

[0017] In one possible implementation, the collaborative scheduling node includes an intent recognition node and a supervisory agent scheduling node; correspondingly, the execution iterative scheduling process includes: parsing the third global state to obtain a decision result; if the decision result is a single-threaded mode, then based on the third global state, determining a first functional node using a state-driven routing algorithm; executing the node logic corresponding to the first functional node to obtain an execution result; generating a final response for the user request based on the execution result; if the decision result is a coroutine concurrent mode, then based on the third global state, determining a supervisory agent scheduling node using a state-driven routing algorithm; using the supervisory agent scheduling node to collaboratively schedule and execute multiple target nodes for the task content in the third global state to generate a final response for the user request; if the decision result is a distributed scheduling mode, then based on the third global state, determining a supervisory agent scheduling node using a state-driven routing algorithm; using the supervisory agent scheduling node to collaboratively schedule and execute multiple target nodes for the task content in the third global state, and initiating a quality monitoring mechanism to generate a final response for the user request.

[0018] In one possible implementation, the functional nodes correspond one-to-one with functional sub-intelligent agents, and the business nodes correspond one-to-one with business sub-intelligent agents. Accordingly, the step of coordinating and executing the task content in the third global state across multiple target nodes to generate a final response to the user request includes: parsing and decomposing the task content in the third global state to obtain multiple sub-tasks; matching the multiple sub-tasks with corresponding multiple target sub-intelligent agents to obtain a collaborative execution plan; allocating the multiple sub-tasks to multiple target nodes corresponding to the multiple target sub-intelligent agents according to the collaborative execution plan; executing the node logic corresponding to the multiple target nodes to obtain multiple execution results; and generating a final response to the user request based on the multiple execution results.

[0019] In one possible implementation, matching multiple target sub-agents corresponding to the multiple sub-tasks to obtain a collaborative execution plan includes: calculating the semantic similarity between the multiple sub-tasks and the functional specializations of each candidate sub-agent; obtaining real-time load data of each candidate sub-agent; determining multiple target sub-agents corresponding to the multiple sub-tasks from the candidate sub-agents based on the semantic similarity and the real-time load data; and obtaining an allocation plan based on the correspondence between the multiple target sub-agents and multiple target nodes.

[0020] In one possible implementation, the node logic corresponding to the business node includes: acquiring key information required for the business process; calculating the information completeness of the current information in the current global state based on the key information; if the information completeness is lower than a preset completeness threshold, generating an incomplete information flag and updating the incomplete information flag to the global state to trigger an information collection and cyclic completion mechanism; if the information completeness is not lower than the preset completeness threshold, executing the business process to obtain an execution result.

[0021] In one possible implementation, the information collection and cyclic completion mechanism includes: prioritizing the missing key information in the global state containing the incomplete information flag, and updating the global state according to the prioritization result to obtain a fourth global state; determining a second functional node based on the fourth global state using a state-driven routing algorithm; generating natural language prompts sequentially by the second functional node according to the priority of the missing information items to guide the user to complete the information; receiving the user's response to the natural language prompts and updating the fourth global state with the response information to obtain a fifth global state; and returning the fifth global state to the business node to repeatedly execute the step "calculating the information completeness of the current information in the current global state" until the information completeness reaches the completeness threshold.

[0022] In one possible implementation, the iterative scheduling process further includes: real-time monitoring of the current global state, and determining whether the human intervention triggering conditions are met based on the current global state. The triggering conditions include one or more of the following: the complexity score exceeds a preset extremely high threshold, the user request explicitly requests a transfer to human intervention, or the cumulative number of execution errors by the sub-agent exceeds a limit. When the human intervention triggering conditions are met, the full data of the current global state and the complete dialogue history context are pushed to the human agent's workbench in real time. The processing result from the workbench is received, and the global state is updated based on the processing result to continue or terminate the iterative scheduling process.

[0023] In one possible implementation, after updating the global state based on the execution result, the method further includes: determining whether the target node is a preset key node; if the target node is a preset key node, serializing the current global state completely and saving it as a checkpoint; saving the checkpoint to a local cache; asynchronously persisting the checkpoint from the local cache to a distributed cluster; when the iterative scheduling process is detected to be interrupted and restarted, restoring the latest checkpoint from the distributed cluster; if the distributed cluster is unavailable, loading the latest checkpoint from a preset archive database.

[0024] Secondly, this application provides a business process processing apparatus, comprising:

[0025] The construction module is used to construct a dynamic graph structure according to a preset business process. The dynamic graph structure contains multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes.

[0026] The receiving module is used to receive user requests sent by the user terminal and initialize the global state according to the user requests;

[0027] The execution module is used to execute an iterative scheduling process in the dynamic graph structure: based on the current global state, determine the target node to be executed through a state-driven routing algorithm; execute the node logic corresponding to the target node to obtain the execution result; update the global state according to the execution result; and repeatedly execute the iterative scheduling process until a final response to the user request is generated.

[0028] The sending module is used to send the final response to the user terminal.

[0029] Thirdly, this application provides a business process processing device, including: a memory and a processor;

[0030] The memory stores computer-executed instructions;

[0031] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0032] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0033] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0034] The business process processing method, apparatus, equipment, storage medium, and program products provided in this application flexibly match different business processes through a dynamic graph structure, avoiding process rigidity; through a state-driven routing algorithm, the target node is accurately located based on the real-time updated global state, reducing invalid execution steps; the iterative scheduling process can gradually improve the processing of user needs, ensuring that the final response meets user requirements, and improving the overall flexibility, accuracy, and efficiency of business process processing. Attached Figure Description

[0035] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0036] Figure 1 A schematic diagram illustrating a business process processing method provided in an embodiment of this application;

[0037] Figure 2 A flowchart illustrating a business process processing method provided in one embodiment of this application;

[0038] Figure 3 A schematic diagram illustrating a business process processing method provided in another embodiment of this application;

[0039] Figure 4 This is a schematic diagram of the structure of the business process processing apparatus provided in the embodiments of this application;

[0040] Figure 5 This is a schematic diagram of the structure of the business process processing device provided in the embodiments of this application.

[0041] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0042] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0043] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0044] Currently, intelligent customer service systems typically process user requests through linear processes or predefined static rules. However, existing technologies, relying on linear processes or predefined static rules, struggle to handle the demands of complex scenarios, resulting in low processing efficiency and insufficient accuracy.

[0045] The business process processing method provided in this application introduces a dynamic graph structure, abstracting the business process into dynamically branchable graph nodes, and combines global state to achieve cross-node data sharing, enabling flexible selection of execution paths. The accuracy of target node selection is optimized through a state-driven routing algorithm based on global state, thereby improving the accuracy and efficiency of business process processing.

[0046] Figure 1 This is a schematic diagram of a scenario for the business process processing method provided in the embodiments of this application, such as... Figure 1 As shown, the specific application scenarios of this application include: user terminal 101 and service device 102.

[0047] The service device 102 can be a server. Optionally, it can be a single server or a cluster of multiple servers.

[0048] Among them, the user terminal 101 can be a mobile phone or a computer.

[0049] Specifically, the user terminal 101 sends a user request to the service device 102; the service device 102 processes the user request to generate a final response to the user request; and the service device 102 sends the final response to the user terminal 101.

[0050] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0051] Figure 2 This is a flowchart illustrating a business process processing method provided in one embodiment of this application. The execution entity in this embodiment can be... Figure 1The service device 102 shown can also be other computer devices, and this embodiment does not impose any particular limitations on it. Figure 2 As shown, the method includes:

[0052] S201: Construct a dynamic graph structure based on the preset business process. The dynamic graph structure contains multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes.

[0053] Optionally, functional nodes correspond one-to-one with functional sub-intelligent agents, business nodes correspond one-to-one with business sub-intelligent agents, and collaborative scheduling nodes include intent recognition nodes and supervisory agent scheduling nodes.

[0054] Optionally, a dynamic graph structure can be constructed based on a preset business process, including: constructing a dynamic graph structure using the LangGraph framework.

[0055] S202: Receive user requests sent by the client and initialize the global state according to the user requests.

[0056] Optionally, the global state is initialized based on the user request, including: parsing the user request to obtain the user identifier and input content; generating a unique session identifier based on the user request; and initializing the global state based on the session identifier, user identifier, and input content.

[0057] Optionally, the global state is initialized based on the session identifier, user identifier, and input content, including: saving the session identifier and user identifier to the global state; initializing the dialogue history to an empty array; setting the task progress to the starting state; obtaining the initialization timestamp; and saving the dialogue history, task progress, and timestamp to the global state.

[0058] The global state includes: session identifier, user identifier, dialogue history, task progress, and timestamp.

[0059] S203: In the dynamic graph structure, execute the iterative scheduling process: Based on the current global state, determine the target node to be executed through the state-driven routing algorithm; execute the node logic corresponding to the target node to obtain the execution result; update the global state according to the execution result; repeat the iterative scheduling process until the final response to the user request is generated.

[0060] Optionally, based on the current global state, a state-driven routing algorithm is used to determine the target node to be executed, including: performing route matching based on preset business rules according to the current information in the global state to obtain a first matching result; if the first matching result is not empty, then determining the first matching result as the target node to be executed; if the first matching result is empty, then obtaining the current user request based on the current information in the global state; calculating the correlation score between each candidate node and the current user request; and determining the target node to be executed from each candidate node based on the correlation score.

[0061] By prioritizing rules, we ensure that regular and clear user requests can be quickly and accurately matched to target nodes, in accordance with business specifications. For fuzzy and complex requests, we use correlation scoring to filter the optimal node, improving the flexibility and accuracy of routing and enhancing the accuracy and efficiency of the overall business process.

[0062] Optionally, based on the current information in the global state, route matching is performed according to preset business rules to obtain a first matching result, including: extracting the user identifier and dialogue history from the global state; obtaining the business type based on the dialogue history; performing matching processing based on the business type and preset business rules; and determining the target node according to the rules if a matching rule exists.

[0063] Optionally, based on the association score, the target node to be executed is determined from each candidate node, including: determining the candidate node with the highest association score as the target node to be executed; when multiple candidate nodes have the highest and the same association score, a round-robin mechanism is used to determine the target node to be executed from the multiple candidate nodes.

[0064] By using a round-robin mechanism, scheduling conflicts when multiple optimal nodes appear are avoided, node load is balanced, and the reliability, adaptability, and efficiency of the state-driven routing algorithm are improved as a whole, thereby improving the accuracy and efficiency of the overall business process.

[0065] Optionally, the association score between each candidate node and the current user request is calculated, including: calculating the semantic matching degree between each candidate node and the current user request; calculating the historical execution success rate of each candidate node; and performing weighted processing based on the semantic matching degree and the historical execution success rate to obtain the association score between each candidate node and the current user request.

[0066] By using semantic matching, we ensure that candidate nodes are compatible with users' ambiguous requests and avoid routing to irrelevant nodes. By combining historical execution success rate factors, we improve the reliability of routing decisions, prioritizing nodes that have been more efficient and less prone to errors in handling similar requests in the past, thus reducing execution errors. This overcomes the limitations of single-rule routing, allowing even ambiguous and complex user requests to be accurately routed to the optimal node, ensuring the continuous progress of the iterative scheduling process and improving the flexibility and accuracy of customer service decisions in business process processing.

[0067] Optionally, before calculating the semantic matching degree between each candidate node and the current user request, the method further includes: filtering out multiple candidate nodes that are currently available and related to the user request in the dynamic graph structure.

[0068] S204: Send the final response to the client.

[0069] The business process processing method provided in this application embodiment flexibly matches different business processes through a dynamic graph structure, avoiding process rigidity; it accurately locates target nodes based on real-time updated global states through a state-driven routing algorithm, reducing invalid execution steps; and the iterative scheduling process can gradually improve the processing of user needs, ensuring that the final response meets user requirements, thereby improving the overall flexibility, accuracy, and efficiency of business process processing.

[0070] In one embodiment of this application, based on the above embodiments, step S203, which involves executing the iterative scheduling process, is further provided in another way, as detailed below:

[0071] SA-a: Based on the first global state, the first cooperative scheduling node is determined through a state-driven routing algorithm; the first global state is the initialized global state.

[0072] SA-b: The first collaborative scheduling node performs in-depth analysis of user requests to obtain intent classification labels and confidence levels.

[0073] Optionally, a deep analysis of the user request is performed to obtain intent classification labels and confidence scores, including: parsing the user request to obtain input content; and using an integrated language model to perform intent recognition and classification processing on the input content to obtain intent classification labels and corresponding confidence scores.

[0074] SA-c: The first collaborative scheduling node performs feature extraction processing on user requests to obtain multi-dimensional feature data.

[0075] Optionally, feature extraction processing is performed on the user request to obtain multi-dimensional feature data, including: performing feature extraction processing on the user request to obtain the length of the input text, the current interaction round of the dialogue, and the risk level of the business itself; and determining the multi-dimensional feature data based on the length of the input text, the current interaction round of the dialogue, and the risk level of the business itself.

[0076] SA-d: The complexity score is calculated by the first collaborative scheduling node based on multi-dimensional feature data.

[0077] Optionally, a complexity score is calculated based on multi-dimensional feature data, including: weighting the length of the input text, the number of interaction rounds in the current dialogue, and the risk level of the business itself using preset weights to obtain a complexity score.

[0078] SA-e: The first global state is updated to the second global state by the first collaborative scheduling node based on the intent classification label, confidence level, and complexity score.

[0079] The business process processing method provided in this application improves the targeting of routing decisions by quantifying complexity scores.

[0080] In one embodiment of this application, based on the above embodiments, step S203, which involves executing the iterative scheduling process, is further provided in another way, as detailed below:

[0081] SB-a: Based on the second global state, the second cooperative scheduling node is determined through a state-driven routing algorithm.

[0082] SB-b: The decision result is determined by the second collaborative scheduling node based on the complexity score and the preset routing decision function.

[0083] Optionally, determining the decision result based on the complexity score and a preset routing decision function includes: judging the relationship between the complexity score and a preset first complexity threshold and a second complexity threshold; wherein the first complexity threshold is less than the second complexity threshold; if the complexity score is less than the first complexity threshold, then the decision result for the single-threaded mode is determined; if the complexity score is greater than the first complexity threshold and less than the second complexity threshold, then the decision result for the coroutine concurrent mode is determined; if the complexity score is greater than the second complexity threshold, then the decision result for the distributed scheduling mode is determined.

[0084] Routing decisions are determined by comparing complexity scores with complexity thresholds. Different execution modes are selected based on the complexity scores: a single-threaded mode is used for simple requests to avoid the overhead of complex scheduling; a coroutine concurrent mode is used for medium-complexity requests to improve execution efficiency within a single thread; and a distributed scheduling mode is used for high-complexity requests to make full use of multi-node resources and improve processing capacity, thereby achieving efficient processing of requests of different complexities.

[0085] SB-c: Update the second global state to the third global state based on the decision result.

[0086] The business process processing method provided in this application selects an appropriate routing decision based on a complexity score and a preset routing decision function. This can optimize resource utilization efficiency and improve the overall processing capability and efficiency of the business process while ensuring processing quality.

[0087] In one embodiment of this application, based on the above embodiments, step S203, which involves executing the iterative scheduling process, is further provided in another way, as detailed below:

[0088] SC-a: Analyze the third global state to obtain the decision result.

[0089] Optionally, parsing the third global state to obtain the decision result includes: parsing the third global state to obtain multiple field information; and extracting the decision result information from the multiple field information.

[0090] SC-b: If the decision result is a single-threaded mode, then based on the third global state, the first functional node is determined by the state-driven routing algorithm; the node logic corresponding to the first functional node is executed to obtain the execution result; and the final response to the user request is generated based on the execution result.

[0091] SC-c: If the decision result is a coroutine concurrent mode, then based on the third global state, the supervisor agent scheduling node is determined through a state-driven routing algorithm; the supervisor agent scheduling node coordinates and executes the task content in the third global state through multiple target nodes to generate the final response to the user request.

[0092] Optionally, the task content in the third global state is coordinated and executed by multiple target nodes to generate a final response to the user request. This includes: parsing and decomposing the task content in the third global state to obtain multiple sub-tasks; matching multiple target sub-agents to the multiple sub-tasks to obtain a coordinated execution plan; allocating the multiple sub-tasks to multiple target nodes corresponding to the multiple target sub-agents according to the coordinated execution plan; executing the node logic corresponding to the multiple target nodes to obtain multiple execution results; and generating a final response to the user request based on the multiple execution results.

[0093] By breaking down complex tasks into multiple simple subtasks, and then using a combination of semantic similarity and real-time load data, the subtasks are assigned to the most suitable sub-agents, thereby improving the efficiency and quality of subtask execution and ultimately enhancing the overall efficiency of business process handling.

[0094] Optionally, multiple target sub-agents are matched for multiple sub-tasks, including: matching target sub-agents from candidate functional sub-agents and business sub-agents according to the business type and / or task requirements of the sub-tasks.

[0095] Optionally, generating a final response to a user request based on multiple execution results includes: summarizing and sorting the multiple execution results; determining whether the sorted multiple execution results are all standardized results; if the sorted multiple execution results are all standardized results, then filling the multiple execution results into a preset response template to obtain a final response; if the sorted multiple execution results contain non-standardized results, then using a preset language model to generate a natural language description based on the multiple execution results to obtain a final response.

[0096] Optionally, sorting multiple execution results includes: sorting and logically organizing multiple execution results according to a preset sorting rule. The preset sorting rule can be priority, logical order, or relevance.

[0097] Optionally, after generating the final response to the user request based on multiple execution results, the method further includes: performing content verification and sensitive information filtering on the final response; and updating the final response with successful verification and sensitive information filtering to the global state.

[0098] Optionally, multiple target sub-agents are matched for multiple sub-tasks to obtain a collaborative execution plan, including: calculating the semantic similarity between the functional expertise of multiple sub-tasks and each candidate sub-agent; obtaining real-time load data of each candidate sub-agent; determining multiple target sub-agents corresponding to multiple sub-tasks from each candidate sub-agent based on semantic similarity and real-time load data; and obtaining an allocation plan based on the correspondence between multiple target sub-agents and multiple target nodes.

[0099] By matching sub-agents using semantic similarity and real-time load data, it is possible to ensure that sub-tasks are assigned to sub-agents with matching functions and sufficient resources, thereby achieving reasonable allocation of sub-tasks and improving the efficiency of sub-task execution and resource utilization.

[0100] Optionally, real-time load data may include: the number of tasks being processed, the central processing unit (CPU), and memory usage, etc.

[0101] SC-d: If the decision result is a distributed scheduling mode, then based on the third global state, the supervisory agent scheduling node is determined through a state-driven routing algorithm; the supervisory agent scheduling node coordinates and executes the task content in the third global state across multiple target nodes, and initiates a quality monitoring mechanism to generate the final response to the user request.

[0102] Optionally, a quality monitoring mechanism can be activated, including: real-time monitoring of task execution quality; and execution of preset remedial strategies if task execution fails. These preset remedial strategies include: re-execution and reporting to human intervention.

[0103] The business process processing method provided in this application has the following modes: single-threaded mode is suitable for simple requests, ensuring simple and efficient execution; coroutine concurrency mode is suitable for medium-complexity requests, achieving multi-task concurrency within a single thread and improving execution efficiency; and distributed scheduling mode is suitable for high-complexity requests, ensuring the quality and reliability of task execution through multi-node collaboration and quality monitoring mechanisms, thereby achieving efficiency and accuracy in processing business processes for requests of different complexities.

[0104] In one embodiment of this application, based on the above embodiments, node logic corresponding to the business node is also provided, as detailed below:

[0105] SD-a: Obtain key information required for business processes.

[0106] Optionally, key information required for the business process can be obtained, including: obtaining the business process corresponding to the business node; and extracting a list of key information required for the business process.

[0107] SD-b: Calculate the information completeness of the current information in the current global state based on key information.

[0108] Optionally, based on the key information, the information completeness of the current information in the current global state is calculated, including: matching the current information in the statistical global state with the list of key information to obtain the number of matching information; and calculating the proportion of the number of matching information in the list of key information to obtain the information completeness.

[0109] SD-c: If the information completeness is lower than the preset completeness threshold, an incomplete information flag is generated and updated to the global state to trigger the information collection and cyclic completion mechanism.

[0110] Optionally, the information collection and cyclic completion mechanism includes: prioritizing the missing key information in the global state containing the incomplete information flag, and updating the global state according to the ranking result to obtain a fourth global state; determining a second functional node based on the fourth global state using a state-driven routing algorithm; generating natural language prompts sequentially according to the priority of the missing information items through the second functional node to guide the user to complete the information; receiving the user's response to the natural language prompts and updating the fourth global state with the response information to obtain a fifth global state; and returning the fifth global state to the business node to repeat the step "calculating the information completeness of the current information in the current global state" until the information completeness reaches a completeness threshold.

[0111] Optionally, the information collection and loop completion mechanism also includes: terminating information collection and loop completion when the number of loop completion attempts reaches the maximum number of attempts.

[0112] By generating natural language prompts to guide users in supplementing missing key information, missing information can be obtained more efficiently, thereby speeding up the information completion process and ensuring the timely execution of business processes.

[0113] SD-d: If the information completeness is not lower than the preset completeness threshold, then execute the business process to obtain the execution result.

[0114] The business process processing method provided in this application can promptly obtain missing key information by checking the completeness of information and triggering a completion mechanism when the information is incomplete, thereby ensuring the smooth execution of the business process, avoiding process failure due to incomplete information, and improving the reliability of business process execution.

[0115] In one embodiment of this application, based on the above embodiments, a process for determining whether manual intervention is required is further included, as detailed below:

[0116] Real-time monitoring of the current global status, and determination of whether the conditions for human intervention are met based on the current global status. The triggering conditions include one or more of the following: the complexity score exceeds the preset extremely high threshold, the user request explicitly requests human intervention, or the cumulative number of errors executed by the sub-agent exceeds the limit.

[0117] When the conditions for human intervention are met, the current global state data and the complete dialogue history context are pushed to the human agent's workbench in real time.

[0118] Receive the processing results from the workbench and update the global state based on the processing results to continue or terminate the iterative scheduling process.

[0119] The business process processing method provided in this application improves the flexibility of business process processing and the ability to deal with complex situations by setting manual intervention trigger conditions and pushing relevant information to human agents, ensuring that user requests are properly handled.

[0120] In one embodiment of this application, based on the above embodiments, a process for saving checkpoints is further included, as detailed below:

[0121] Determine whether the target node is a preset key node.

[0122] If the target node is a preset critical node, the current global state will be fully serialized and saved as a checkpoint.

[0123] Save checkpoints to the local cache.

[0124] Asynchronously persist checkpoints from the local cache to the distributed cluster.

[0125] Alternatively, the distributed cluster can be a distributed Redis cluster.

[0126] When an interruption in the iterative scheduling process is detected and restarted, the latest checkpoint is restored from the distributed cluster.

[0127] If the distributed cluster is unavailable, the latest checkpoint is loaded from the pre-defined archive database.

[0128] The business process processing method provided in this application, by setting checkpoints and multiple storage mechanisms, can restore the latest checkpoint from the distributed cluster or archive database when the process is interrupted, so that the business process can continue to execute from the point of interruption, thereby improving the reliability and fault tolerance of business process processing.

[0129] One embodiment of this application provides a business process processing system, which includes: a dynamic graph structure workflow module, a global state persistence module, a multi-agent collaboration module, and a demand scheduling and human intervention fusion module.

[0130] Among them, the dynamic graph structure workflow module is used for process topology modeling and state-driven routing; the global state persistence module is used to maintain cross-module state sharing and breakpoint recovery; the multi-agent collaboration module is used to realize dynamic task allocation and function reuse; and the demand scheduling and human intervention integration module is used to adapt to the complexity of the demand and form a human-machine collaboration closed loop.

[0131] Optionally, the dynamic graph structure workflow module, built on the LangGraph framework, provides flexible graph structure modeling and execution capabilities for customer service business processes, specifically including:

[0132] (a) Node Layer Definition and Standardized Encapsulation. A three-layer node architecture of "Function-Business-Collaboration" is adopted. Functional nodes correspond one-to-one with functional sub-agents, defining input and output constraints through JSON data patterns and deploying execution logic in a containerized image manner to ensure environment isolation. Business nodes correspond to business sub-agents, internally using a finite state machine model to manage state transitions such as "initial-execution-end-exception," and have built-in retry and other exception handling mechanisms. The collaborative scheduling nodes include intent recognition nodes and supervisor agent scheduling nodes. The intent recognition node is used to output high-confidence intent classification results based on a fine-tuned pre-trained language model; the supervisor agent scheduling node is responsible for maintaining the agent registry and real-time state information.

[0133] (b) State-driven routing algorithm. A hybrid strategy of "rule priority + probability compensation" is adopted. The rule layer performs precise matching based on preset business rules to achieve efficient routing. When a rule is not matched, the probability compensation layer calculates the correlation score of each candidate node and the current request, as well as the historical success rate of each candidate node, and makes routing decisions based on the correlation score to determine the target node to be executed. At the same time, a conflict handling mechanism such as round-robin is introduced to avoid single point overload.

[0134] (c) Looping Completion Mechanism. This mechanism quantitatively assesses the completeness of information in the global state and automatically triggers interactive completion between the information collection node and the user when information is missing. The completion process is based on the information weight priority and is cyclically verified until the information meets the completeness threshold or reaches the maximum number of attempts, ensuring that the business process can proceed smoothly.

[0135] Optionally, a global state persistence module is used to address state consistency, high availability, and fault recovery issues in a distributed environment, specifically including:

[0136] (a) Structured modeling of global state. A strongly typed approach is used to define the global state, which includes key information such as user identifier, dialogue history, task progress, cross-system data, complexity score, information completeness, and checkpoint timestamp. The global state specifies the data type and value constraints for each field, providing a unified and standardized state context for the system.

[0137] (b) Dual Storage Collaboration and Resume from Breakpoint. A dual storage architecture combining local high-speed caching and a distributed Redis cluster is adopted. During state updates, the local cache is updated first in a transactional manner through a checkpoint mechanism, and then asynchronously persisted to the distributed Redis cluster. When the system restarts abnormally, the state can be quickly restored from the Redis cluster based on the session identifier; if the Redis cluster is unavailable, the most recent backup can be loaded from a pre-set archive database storage such as MySQL (My Structured Query Language), thereby enabling the task to resume from breakpoint and ensuring business continuity.

[0138] The multi-agent collaboration module, through a hierarchical collaboration architecture, enables flexible task allocation and efficient execution, specifically including:

[0139] (a) Two-dimensional decision-making model of the supervisor agent. The supervisor agent scheduling node is responsible for the allocation decision of sub-tasks, and its decision-making is mainly based on two dimensions: functional matching degree and system load. Functional matching degree is evaluated by calculating the semantic similarity between user requirements and the functional expertise of each sub-agent; system load is comprehensively calculated by monitoring indicators such as CPU, memory and queue length in real time. The decision rule prioritizes sub-agents with high matching degree and low load, or selects the available agent with the lowest load under specific conditions, so as to achieve efficient task processing under load balancing.

[0140] (b) Standardized Design of Functional Sub-Agents. Functional sub-agents focus on the execution of atomic operations, such as authentication, information retrieval, or network diagnostics. All functional sub-agents provide standardized RESTful (Representational State Transfer) APIs (Application Programming Interfaces), and adhere to the OpenAPI specification to clearly define request methods, URLs (Uniform Resource Locators), request / response body structures, and performance metrics (such as response latency), ensuring the standardization and reusability of the interfaces.

[0141] (c) Business-oriented sub-agent process architecture. Business-oriented sub-agents are responsible for handling complex business processes involving multiple steps (such as fault reporting). Internally, they use a process definition language (such as BPEL (Business Process Execution Language)) to describe node sequences, transition conditions, and exception handling logic. Business-oriented sub-agents reuse the capabilities of functional sub-agents through dependency injection, maintain their own process state, and keep synchronized with the global state through an event mechanism.

[0142] Optionally, a demand scheduling and human intervention integration module is used to achieve dynamic resource scheduling and seamless human-machine collaboration based on demand complexity, specifically including:

[0143] (a) Quantitative Model for Demand Complexity. A quantitative complexity score for the current user demand is calculated by weighted summation, integrating standardized scores from multiple dimensions such as text length, number of interaction rounds, and business risk level. The weights of each dimension are determined based on historical data statistical analysis to ensure the accuracy of the complexity assessment.

[0144] (b) Dynamic scheduling strategy for intelligent agents. Based on the calculated complexity score, the scheduling strategy and resource allocation are dynamically adjusted. For low-complexity requirements, a single-threaded mode is used for rapid response; for medium-complexity requirements, a coroutine concurrent mode is enabled and a supervisor intelligent agent coordinating node is introduced for coordination; for high-complexity requirements, a distributed scheduling mode is adopted, mobilizing more intelligent agent resources and introducing a quality monitoring mechanism to cope with complex scenarios.

[0145] (c) Human Intervention and Model Optimization Mechanism. The system pre-sets various human intervention trigger conditions, such as extremely high complexity, user requests explicitly requiring human or intelligent agent error rates to exceed limits, etc. When the conditions are met, the full data of the current global state is synchronized to the human agent's workbench in real time to obtain the human processing results. The human processing results are not only used to update the state and complete the service loop, but also used as supervision samples to optimize relevant model parameters in the system, achieving continuous learning.

[0146] Figure 3 This is a schematic diagram illustrating a business process processing method provided in another embodiment of this application. For example... Figure 3 As shown, the method includes:

[0147] Step 1: User request access and state initialization.

[0148] Optionally, the system receives user requests through a standardized interface, parsing the user identifier and input content within the request. Subsequently, it initializes a global state object: generating a unique session identifier, recording the user identifier, initializing the dialogue history as an empty array, setting the current task progress to the starting state, and recording the initialization timestamp. Simultaneously, a basic process topology structure containing start, end, and core functional nodes is established in the dynamic graph engine.

[0149] Step 2: Intent identification and requirement complexity assessment.

[0150] Optionally, the system invokes an integrated language model to perform deep analysis of user input, identifying its core intent (such as querying, processing, or complaining) and outputting corresponding intent classification labels and confidence levels. Simultaneously, based on multiple dimensions of feature data, including the length of the input text, the number of interaction rounds in the current dialogue, and the risk level of the business itself, a quantitative requirement complexity score is comprehensively calculated. The intent classification labels and complexity scores are then updated in the global state.

[0151] Step 3: Dynamic routing decision based on complexity.

[0152] Optionally, a routing decision function is used to automatically match the target node to be executed next based on the complexity score in the global state. The decision logic adopts a hierarchical strategy, including judging the complexity level based on the complexity score: if the complexity score is less than the first complexity threshold, it is low complexity, and low complexity tasks adopt a single-threaded mode: routing to a single functional node (functional sub-agent) for fast processing; if the complexity score is greater than the first complexity threshold but less than the second complexity threshold, it is medium complexity, and medium complexity tasks trigger a coroutine concurrent mode; if the complexity score is greater than the second complexity threshold, it is high complexity, and high complexity tasks adopt a distributed scheduling mode. The decision result drives the graph engine to jump to the corresponding node and updates the task progress in the state. Both the coroutine concurrent mode and the distributed scheduling mode employ multi-agent collaborative processing; specifically, in the coroutine concurrent mode, the supervisor agent schedules the node to allocate tasks to a first preset number of functional nodes (functional sub-agents); in the distributed scheduling mode, the supervisor agent schedules the node to allocate tasks to a second preset number of functional nodes (functional sub-agents) and initiates a quality monitoring mechanism.

[0153] The first preset quantity is less than the second preset quantity.

[0154] Optionally, the first complexity threshold can be 0.3, and the second complexity threshold can be 0.7.

[0155] Step 4: Triggering complex tasks and multi-agent collaboration.

[0156] When the complexity score is determined to be high or medium complexity, the system automatically triggers the intervention of the supervisor agent scheduling node. The supervisor agent scheduling node is responsible for decomposing the complex task, analyzing the matching degree between the subtasks and the capabilities of each specialized agent, and comprehensively considering the real-time load of each agent to formulate a collaborative execution plan.

[0157] Step 5: Subtask allocation and agent execution.

[0158] The supervisory agent scheduling node assigns the decomposed subtasks to multiple functional or business-oriented sub-agents. Each sub-agent receives standardized instructions and executes its specialized atomic operations (such as authentication or inventory lookup) or encapsulated business processes. Business-oriented sub-agents can also sequentially call multiple functional sub-agents according to predefined processes, achieving interface reuse.

[0159] Step Six: Completeness check of business node information.

[0160] Optionally, if the decomposed subtasks are assigned to business-oriented sub-agents, when the process enters the business-oriented node (e.g., handling business) corresponding to the business-oriented sub-agent, the business-oriented node first checks whether the key information required for execution (e.g., user identity, business parameters, etc.) has been completely collected in the global state. The system will calculate the completeness of the current information. If the completeness is lower than the preset completeness threshold, an incomplete information flag will be generated, and an information collection and cyclic completion mechanism will be triggered.

[0161] Optionally, the preset completeness threshold can be 90%.

[0162] Step 7: Information collection and cyclical completion mechanism.

[0163] Optionally, if the information completeness is insufficient, the process will jump to the information collection node. The information collection node generates a prompt message to guide the user to supplement the missing information based on its priority (weight), thus collecting the missing fields. After obtaining the user's reply, the reply is parsed and updated to the global state, and the information completeness is recalculated. This process is repeated until the information completeness meets the preset completeness threshold. Then, the process jumps back to the original business node to continue executing the node logic and obtain the result.

[0164] Among them, the information collection node is a functional node.

[0165] Step 8: Result aggregation and natural language response generation.

[0166] Each sub-agent returns its execution result to the supervisor agent scheduling node. The supervisor agent scheduling node summarizes, sorts, and integrates the multiple execution results. For standardized results, they are filled into a preset response template to obtain the final response; for non-standardized content, a language model is invoked to generate a fluent natural language description to obtain the final response. After content validation and sensitive information filtering are completed, the final response is updated to the global state.

[0167] Step 9: Save and synchronize status checkpoints.

[0168] At pre-defined critical nodes (such as after a task is completed), the system automatically saves the current global state as a checkpoint. This process first updates the local cache to ensure fast reading, and then asynchronously synchronizes the entire state to the distributed cluster for persistence, ensuring state recoverability and consistency in the cluster environment.

[0169] Step 10: Human intervention for judgment and seamless switching.

[0170] The system monitors in real time whether the conditions for human intervention are met (e.g., extremely high complexity, user requests explicitly requiring human or AI agent processing exceeding the error limit, etc.). If the conditions for human intervention are met, the system will push the current complete dialogue context and global state snapshot to the human customer service agent's workbench in real time via a persistent connection. It will receive the human processing results returned from the human customer service agent's workbench, update the global state dialogue history, and optimize the model. If the conditions for human intervention are not met, steps two through nine will be executed to output the final response to the client.

[0171] Step 11: Session persistence and data archiving.

[0172] After the entire session ends, the system will perform final data persistence processing, including: saving session data in high-performance storage for fast querying, and periodically saving historical session data to an archive database for long-term retention, compliance auditing, and subsequent model optimization and analysis.

[0173] The archive database is a large-capacity database, and the high-performance storage can be a distributed cluster.

[0174] The business process processing method provided in this application performs in-depth analysis of user requests and improves the targeting and overall processing capability of business processes by calculating the appropriate execution mode through computational complexity scoring. It decomposes complexity into sub-tasks and matches them with corresponding sub-agents, allocating tasks by comprehensively considering semantic similarity and real-time load data, thereby improving the execution efficiency of sub-tasks and the utilization rate of system resources. Through information completeness checks and manual intervention mechanisms, it ensures the reliability of business process execution and the ability to handle complex and special situations. Furthermore, by saving checkpoints and employing a multi-storage mechanism, it enhances the fault tolerance and recovery capabilities of the business process. Overall, it improves the flexibility, accuracy, and efficiency of business process processing.

[0175] Figure 4 This is a schematic diagram of the structure of the business process processing device provided in the embodiments of this application, such as... Figure 4 As shown, the business process processing device provided in this embodiment includes: a construction module 401, a receiving module 402, an execution module 403, and a sending module 404.

[0176] The construction module 401 is used to construct a dynamic graph structure according to a preset business process. The dynamic graph structure contains multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes.

[0177] The receiving module 402 is used to receive user requests sent by the user terminal and initialize the global state according to the user requests;

[0178] The execution module 403 is used to execute an iterative scheduling process in a dynamic graph structure: based on the current global state, it determines the target node to be executed through a state-driven routing algorithm; executes the node logic corresponding to the target node to obtain the execution result; updates the global state according to the execution result; and repeats the iterative scheduling process until the final response to the user request is generated.

[0179] The sending module 404 is used to send the final response to the user terminal.

[0180] In one possible implementation, the execution module 403 is specifically used to: perform route matching based on preset business rules according to the current information in the global state to obtain a first matching result; if the first matching result is not empty, determine the first matching result as the target node to be executed; if the first matching result is empty, obtain the current user request based on the current information in the global state; calculate the correlation score between each candidate node and the current user request; and determine the target node to be executed from each candidate node based on the correlation score.

[0181] In one possible implementation, the execution module 403, when calculating the association score between each candidate node and the current user request, is specifically used to: calculate the semantic matching degree between each candidate node and the current user request; calculate the historical execution success rate of each candidate node; and perform weighted processing based on the semantic matching degree and the historical execution success rate to obtain the association score between each candidate node and the current user request.

[0182] In one possible implementation, the execution module 403 determines the target node to be executed from the candidate nodes based on the association score. Specifically, it is used to: determine the candidate node with the highest association score as the target node to be executed; when multiple candidate nodes have the highest and the same association score, a round-robin mechanism is used to determine the target node to be executed from the multiple candidate nodes.

[0183] In one possible implementation, execution module 403 is specifically used for: determining a first cooperative scheduling node based on a first global state using a state-driven routing algorithm; the first global state is an initialized global state; performing in-depth analysis of user requests through the first cooperative scheduling node to obtain intent classification labels and confidence levels; performing feature extraction processing on user requests through the first cooperative scheduling node to obtain multi-dimensional feature data; calculating a complexity score through the first cooperative scheduling node based on the multi-dimensional feature data; and updating the first global state to a second global state through the first cooperative scheduling node based on the intent classification labels, confidence levels, and complexity scores.

[0184] In one possible implementation, the execution module 403 is specifically used to: determine a second cooperative scheduling node based on the second global state using a state-driven routing algorithm; determine a decision result using the second cooperative scheduling node based on a complexity score and a preset routing decision function; and update the second global state to a third global state based on the decision result.

[0185] In one possible implementation, the execution module 403, when determining the decision result based on the complexity score and a preset routing decision function, specifically performs the following: determining the relationship between the complexity score and a preset first complexity threshold and a second complexity threshold; wherein the first complexity threshold is less than the second complexity threshold; if the complexity score is less than the first complexity threshold, then the decision result for the single-threaded mode is determined; if the complexity score is greater than the first complexity threshold and less than the second complexity threshold, then the decision result for the coroutine concurrent mode is determined; if the complexity score is greater than the second complexity threshold, then the decision result for the distributed scheduling mode is determined.

[0186] In one possible implementation, the collaborative scheduling node includes an intent recognition node and a supervisory agent scheduling node. Accordingly, the execution module 403 is specifically used for: parsing the third global state to obtain a decision result; if the decision result is a single-threaded mode, then determining the first functional node based on the third global state using a state-driven routing algorithm; executing the node logic corresponding to the first functional node to obtain an execution result; generating a final response to the user request based on the execution result; if the decision result is a coroutine concurrent mode, then determining the supervisory agent scheduling node based on the third global state using a state-driven routing algorithm; using the supervisory agent scheduling node to collaboratively schedule and execute multiple target nodes for the task content in the third global state to generate a final response to the user request; if the decision result is a distributed scheduling mode, then determining the supervisory agent scheduling node based on the third global state using a state-driven routing algorithm; using the supervisory agent scheduling node to collaboratively schedule and execute multiple target nodes for the task content in the third global state, and initiating a quality monitoring mechanism to generate a final response to the user request.

[0187] In one possible implementation, functional nodes correspond one-to-one with functional sub-intelligent agents, and business nodes correspond one-to-one with business sub-intelligent agents. Accordingly, when the execution module 403 performs coordinated scheduling and execution of multiple target nodes on the task content in the third global state to generate a final response to the user request, it specifically includes: parsing and decomposing the task content in the third global state to obtain multiple sub-tasks; matching multiple target sub-intelligent agents to the multiple sub-tasks to obtain a coordinated execution plan; allocating the multiple sub-tasks to multiple target nodes corresponding to the multiple target sub-intelligent agents according to the coordinated execution plan; executing the node logic corresponding to the multiple target nodes to obtain multiple execution results; and generating a final response to the user request based on the multiple execution results.

[0188] In one possible implementation, the execution module 403 matches multiple target sub-agents for multiple sub-tasks to obtain a collaborative execution plan. Specifically, it is used to: calculate the semantic similarity between the functional specializations of multiple sub-tasks and each candidate sub-agent; obtain the real-time load data of each candidate sub-agent; determine the multiple target sub-agents corresponding to the multiple sub-tasks from each candidate sub-agent based on the semantic similarity and the real-time load data; and obtain an allocation plan based on the correspondence between the multiple target sub-agents and multiple target nodes.

[0189] In one possible implementation, the node logic corresponding to the business node includes: acquiring key information required for the business process; calculating the information completeness of the current information in the current global state based on the key information; if the information completeness is lower than a preset completeness threshold, generating an incomplete information flag and updating the incomplete information flag to the global state to trigger an information collection and cyclic completion mechanism; if the information completeness is not lower than the preset completeness threshold, executing the business process to obtain the execution result.

[0190] In one possible implementation, the information collection and cyclic completion mechanism includes: prioritizing the missing key information in the global state containing the incomplete information flag, and updating the global state according to the prioritization result to obtain a fourth global state; determining a second functional node based on the fourth global state using a state-driven routing algorithm; generating natural language prompts sequentially according to the priority of the missing information items through the second functional node to guide the user to complete the information; receiving the user's response to the natural language prompts and updating the fourth global state with the response information to obtain a fifth global state; and returning the fifth global state to the business node to repeat the step "calculating the information completeness of the current information in the current global state" until the information completeness reaches a completeness threshold.

[0191] In one possible implementation, the business process processing device further includes:

[0192] The judgment module is used to monitor the current global state in real time and determine whether the conditions for human intervention are met based on the current global state. The triggering conditions include one or more of the following: the complexity score exceeds a preset extremely high threshold, the user request explicitly requests human intervention, or the cumulative number of execution errors by the sub-agent exceeds the limit. When the conditions for human intervention are met, the module pushes the full data of the current global state and the complete dialogue history context to the human agent's workbench in real time. The module receives the processing results from the workbench and updates the global state based on the processing results to continue or terminate the iterative scheduling process.

[0193] In one possible implementation, the business process processing device further includes:

[0194] The save module is used to determine whether the target node is a preset critical node. If the target node is a preset critical node, the current global state is completely serialized and saved as a checkpoint. The checkpoint is saved to the local cache. The checkpoint is asynchronously persisted to the distributed cluster from the local cache. When the iterative scheduling process is restarted after being interrupted, the latest checkpoint is restored from the distributed cluster. If the distributed cluster is unavailable, the latest checkpoint is loaded from the preset archive database.

[0195] The business process processing device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0196] Figure 5 This is a schematic diagram of the structure of a business process processing device provided in an embodiment of this application. Figure 5 As shown, the business process processing device provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0197] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0198] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0199] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0200] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0201] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0202] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0203] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0204] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0205] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0206] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0207] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0208] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0209] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0211] Finally, it should be noted that other embodiments of this application will readily conceive of by those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and alterations may be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A business process processing method, characterized in that, include: A dynamic graph structure is constructed based on a preset business process. The dynamic graph structure contains multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes. Receive user requests sent by the client and initialize the global state according to the user requests; In the dynamic graph structure, an iterative scheduling process is executed: based on the current global state, the target node to be executed is determined through a state-driven routing algorithm; Execute the node logic corresponding to the target node to obtain the execution result; Update the global state based on the execution result; Repeat the iterative scheduling process until a final response to the user request is generated; The final response is sent to the user terminal.

2. The method according to claim 1, characterized in that, The process of determining the target node to be executed based on the current global state using a state-driven routing algorithm includes: Based on the current information in the global state, route matching is performed according to preset business rules to obtain the first matching result; If the first matching result is not empty, then the first matching result is determined to be the target node to be executed; If the first matching result is empty, then the current user request is obtained based on the current information in the global state; the association score between each candidate node and the current user request is calculated; and the target node to be executed is determined from each candidate node based on the association score.

3. The method according to claim 2, characterized in that, The calculation of the correlation score between each candidate node and the current user request includes: Calculate the semantic matching degree between each candidate node and the current user request; Calculate the historical execution success rate of each candidate node; The semantic matching degree and the historical execution success rate are weighted to obtain the association score between each candidate node and the current user request.

4. The method according to claim 2, characterized in that, The step of determining the target node to be executed from the candidate nodes based on the association score includes: The candidate node with the highest association score is selected as the target node to be executed. When multiple candidate nodes have the highest and the same association score, a round-robin mechanism is used to determine the target node to be executed from the multiple candidate nodes.

5. The method according to claim 1, characterized in that, The execution iterative scheduling process includes: Based on the first global state, the first cooperative scheduling node is determined by a state-driven routing algorithm; the first global state is the initialized global state. The first collaborative scheduling node performs in-depth analysis on the user request to obtain intent classification labels and confidence levels. The user request is processed by the first collaborative scheduling node to obtain multi-dimensional feature data. The complexity score is calculated by the first collaborative scheduling node based on the multi-dimensional feature data; The first collaborative scheduling node updates the first global state to the second global state based on the intent classification label, confidence level, and complexity score.

6. The method according to claim 5, characterized in that, The execution iterative scheduling process includes: Based on the second global state, the second cooperative scheduling node is determined by a state-driven routing algorithm; The second collaborative scheduling node determines the decision result based on the complexity score and the preset routing decision function. The second global state is updated to the third global state based on the decision result.

7. The method according to claim 6, characterized in that, The step of determining the decision result based on the complexity score and the preset routing decision function includes: Determine the relationship between the complexity score and a preset first complexity threshold and a second complexity threshold; wherein the first complexity threshold is less than the second complexity threshold; If the complexity score is less than the first complexity threshold, then the decision result of the single-threaded mode is determined; If the complexity score is greater than the first complexity threshold and less than the second complexity threshold, then the decision result of the coroutine concurrency mode is determined. If the complexity score is greater than the second complexity threshold, then the decision result of the distributed scheduling mode is determined.

8. The method according to claim 7, characterized in that, The collaborative scheduling node includes an intent recognition node and a supervisory agent scheduling node; Accordingly, the execution iterative scheduling process includes: The third global state is analyzed to obtain the decision result; If the decision result is a single-threaded mode, then based on the third global state, a first functional node is determined through a state-driven routing algorithm; the node logic corresponding to the first functional node is executed to obtain the execution result; and a final response to the user request is generated based on the execution result. If the decision result is a coroutine concurrent mode, then based on the third global state, a state-driven routing algorithm is used to determine the supervisor agent scheduling node; the supervisor agent scheduling node is used to coordinate and execute the task content in the third global state across multiple target nodes to generate a final response to the user request. If the decision result is a distributed scheduling mode, then based on the third global state, a state-driven routing algorithm is used to determine the supervisor agent scheduling node; the supervisor agent scheduling node is used to coordinate and execute the task content in the third global state across multiple target nodes, and a quality monitoring mechanism is initiated to generate a final response to the user request.

9. The method according to claim 8, characterized in that, Each functional node corresponds one-to-one with a functional sub-agent, and each business node corresponds one-to-one with a business sub-agent. Accordingly, the coordinated scheduling and execution of the task content in the third global state across multiple target nodes to generate a final response to the user request includes: The task content in the third global state is parsed and decomposed to obtain multiple sub-tasks; Multiple target sub-agents are matched with the multiple sub-tasks to obtain a collaborative execution plan; According to the collaborative execution plan, the multiple sub-tasks are assigned to the multiple target nodes corresponding to the multiple target sub-agents; Execute the node logic corresponding to the multiple target nodes to obtain multiple execution results; A final response to the user's request is generated based on the multiple execution results.

10. The method according to claim 9, characterized in that, The step of matching multiple target sub-agents corresponding to the multiple sub-tasks to obtain a collaborative execution plan includes: Calculate the semantic similarity between the multiple subtasks and the functional specialties of each candidate sub-agent; Obtain the real-time load data of each candidate sub-agent; Based on the semantic similarity and the real-time load data, multiple target sub-agents corresponding to the multiple sub-tasks are determined from each candidate sub-agent; The allocation plan is obtained based on the correspondence between the multiple target sub-agents and the multiple target nodes.

11. The method according to any one of claims 1 to 10, characterized in that, The node logic corresponding to the business-type node includes: Obtain the key information required for the business process; Based on the key information, calculate the information completeness of the current information in the current global state; If the information completeness is lower than the preset completeness threshold, an incomplete information flag is generated and updated to the global state to trigger the information collection and cyclic completion mechanism. If the information completeness is not lower than the preset completeness threshold, then the business process is executed to obtain the execution result.

12. The method according to claim 11, characterized in that, The information collection and cyclic completion mechanism includes: The missing key information in the global state containing the incomplete information flag is sorted by weight priority, and the global state is updated according to the sorting result to obtain the fourth global state. Based on the fourth global state, the second functional node is determined by a state-driven routing algorithm; The second functional node generates natural language prompts sequentially based on the priority of missing information items to guide the user in completing the missing information. Receive the response information sent by the user to the natural language prompt, and update the response information to the fourth global state to obtain the fifth global state; The fifth global state is returned to the business node to repeat the step "calculate the information completeness of the current information in the current global state" until the information completeness reaches the completeness threshold.

13. The method according to any one of claims 1 to 9, characterized in that, The iterative scheduling process also includes: The system monitors the current global status in real time and determines whether the conditions for human intervention are met based on the current global status. The triggering conditions include one or more of the following: the complexity score exceeds a preset extremely high threshold, the user request explicitly requests human intervention, or the cumulative number of errors by the sub-agent exceeds the limit. When the conditions for human intervention are met, the current global state data and the complete dialogue history context will be pushed to the human agent's workbench in real time. Receive the processing result from the workbench and update the global state based on the processing result to continue or terminate the iterative scheduling process.

14. The method according to any one of claims 1 to 9, characterized in that, After updating the global state based on the execution result, the method further includes: Determine whether the target node is a preset key node; If the target node is a preset key node, then the current global state is fully serialized and saved as a checkpoint; Save the checkpoints to the local cache; The checkpoint is asynchronously persisted to the distributed cluster from the local cache; When the iterative scheduling process is detected to have been interrupted and restarted, the latest checkpoint is restored from the distributed cluster; If the distributed cluster is unavailable, the latest checkpoint is loaded from the preset archive database.

15. A business process processing device, characterized in that, include: The construction module is used to construct a dynamic graph structure according to a preset business process. The dynamic graph structure includes multiple nodes and edges that define the execution relationships between nodes. The nodes include functional nodes, business nodes, and collaborative scheduling nodes. The receiving module is used to receive user requests sent by the user terminal and initialize the global state according to the user requests; The execution module is used to execute an iterative scheduling process in the dynamic graph structure: based on the current global state, it determines the target node to be executed through a state-driven routing algorithm; Execute the node logic corresponding to the target node to obtain the execution result; Update the global state based on the execution result; Repeat the iterative scheduling process until a final response to the user request is generated; The sending module is used to send the final response to the user terminal.

16. A business process processing device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-14.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-14.

18. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-14.

Images