Workflow operation method, device and medium based on multi-agent and human-machine collaboration

Abstract

The present application discloses a workflow operation method, device and medium based on multi-agent and human-machine collaboration, which relates to the field of artificial intelligence. The method includes: splitting business scenarios into workflow tasks and constructing a workflow model; the workflow model is a directed cyclic graph; initializing a human-machine collaborative multi-agent system based on the workflow model; determining the workflow type according to the business request input by the user, and determining the starting agent node according to the workflow type; starting from the starting agent node, the workflow is transferred between each work node based on the workflow model and the human-machine collaborative multi-agent system; during the workflow execution process, human-machine collaboration is performed based on human-machine collaborative edges, business audit is performed based on audit edges and audit agent nodes, and the global state is updated until the process ends when the global state is a global termination state. The present application improves the execution efficiency, execution accuracy and user experience of the workflow.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a workflow operation method, device and medium based on multi-agent and human-machine collaboration. Background Art

[0002] Currently, commercial robots that can perform business through human-machine interaction (such as robots in banks and hospital lobby) implement the mainstream business workflow in the following way: business personnel first sort out the business process, and then developers "hard code" the workflow into the graphical user interface (GUI), turning it into a large application. The more functions, the higher the cost, but the more difficult it is for users to locate and use it. In addition, some telephone ticket booking robots complete the business workflow through a "hard-coded" dialogue process, which prevents users from speaking freely and is too rigid and has a poor experience.

[0003] With the rapid development of large language model technology, native applications of artificial intelligence (AI) have gradually penetrated into people's daily life, study and work. Represented by chat robots, the conversation is free and comfortable, but most of them can only conduct business knowledge base Q&A and cannot actually complete business processing, confirmation and submission. There are also some writing intelligent agent systems that can independently complete actual tasks, mainly chain workflows, but have poor interaction with humans and are mostly used in independent tasks independent of humans. And due to the "hallucination" problem, it is often difficult to actually play a role.

[0004] Therefore, people hope to interact with AI like "human-human collaboration" in reality, and expect robots to have the ability to complete tasks autonomously and be controlled precisely. Generally speaking, the methods for implementing business workflows based on large language model agents mainly include: task decomposition and planning, ReAct framework, tool use, agent reflection, multi-agent collaboration, agent workflow, knowledge retrieval enhancement, and thought chain.

[0005] Although the above methods can better implement task execution using large language models or agents, and can also complete some automated workflow tasks, there are still the following defects when using large language models and multi-agent systems to execute workflow tasks.

[0006] (1) Unreliable generated content: Large language models may experience “hallucinations” when generating content, that is, they may produce inaccurate or false information. This phenomenon may cause decisions in the workflow to be based on incorrect data, affecting the overall task efficiency and reliability.

[0007] (2) Single mode of multi-agent cooperation: The mainstream multi-agent collaboration is a linear acyclic graph chain call mode. The calculation process is usually stateless and each step is independent of each other. It cannot handle workflow tasks that require iterative nesting, feedback loops, and complex decisions, which reduces the effectiveness of the entire system and user experience.

[0008] (3) Difficulty in global workflow control: Workflows are divided into single-interaction processes and multi-round interaction processes. Most mainstream methods are designed to solve single-interaction workflows, but lack global control over multi-rounds, resulting in the inability to effectively manage and coordinate the dependencies and sequences between different steps in complex workflow scenarios.

[0009] (4) Technical barriers and usage complexity: For users who lack relevant technical background, it may be difficult to understand and operate large language models and intelligent agent systems. Such systems usually require expertise in machine learning, data science, or programming, which increases the learning threshold.

[0010] These challenges make the current workflow-oriented multi-agent systems perform poorly in terms of practicality and reliability. Summary of the invention

[0011] The purpose of this application is to provide a workflow operation method, device and medium based on multi-agent and human-computer collaboration, which can improve the execution efficiency, execution accuracy and user experience of the workflow.

[0012] To achieve the above objectives, this application provides the following solutions:

[0013] In a first aspect, the present application provides a workflow operation method based on multi-agent and human-machine collaboration, comprising:

[0014] The business scenario is divided into workflow tasks, and a workflow model is constructed; the workflow model is a directed cyclic graph, including a global state, a start node, a work node, a work edge, a jump condition, a business application program interface, and a constraint condition; wherein the work node includes a work agent node, an audit agent node, and an end node; the work edge includes a global state edge, a human-machine collaboration edge, a common process edge, and an audit edge; the jump condition includes an unconditional jump and a business application program interface jump;

[0015] Based on the workflow model, a human-machine collaborative multi-agent system is initialized; the human-machine collaborative multi-agent system includes a plurality of large language model agents; the plurality of large language model agents serve as start nodes or working nodes in the workflow model;

[0016] Determine the workflow type according to the business request input by the user, and determine the starting agent node according to the workflow type; the workflow type is a single-round workflow or a multi-round workflow; the starting agent node is a starting node or a working agent node;

[0017] Based on the workflow model and the human-machine collaborative multi-agent system, starting from the start agent node, the workflow is transferred and executed between each working node;

[0018] During the workflow execution process, human-machine collaboration is performed based on the human-machine collaboration edge, business audit is performed based on the audit edge and the audit agent node, and the global state is updated until the process ends when the global state is a global termination state.

[0019] In a second aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned workflow operation method based on multi-agent and human-computer collaboration.

[0020] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned workflow operation method based on multi-agent and human-computer collaboration.

[0021] According to the specific embodiments provided by the present application, the present application discloses the following technical effects: The present application provides a workflow operation method, device and medium based on multi-agent and human-machine collaboration. By splitting the business scenario into workflow tasks and constructing a directed cyclic graph, the human-machine collaborative multi-agent system can clearly perform complex tasks such as nesting and looping in the directed cyclic graph, which not only improves the flexibility and adaptability of the human-machine collaborative multi-agent system, but also greatly simplifies the management and execution process of complex workflows, and improves the overall work efficiency. By distinguishing between single-round workflows and multi-round workflows, the problem of global control difficulties is solved, so that the human-machine collaborative multi-agent system can better adapt to complex workflow scenarios and improve the execution accuracy of workflows. During the workflow execution process, human-machine collaboration is carried out based on human-machine collaborative edges, and business audits are carried out based on audit edges and audit agent nodes, which improves the ease of use and user experience of the human-machine collaborative multi-agent system, and also lowers the technical threshold, so that users who lack relevant background knowledge can also operate easily. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 This is an application environment diagram of a workflow operation method based on multi-agent and human-computer collaboration in one embodiment of the present application.

[0024] Figure 2 A flowchart of a workflow operation method based on multi-agent and human-computer collaboration is provided for one embodiment of the present application.

[0025] Figure 3 A schematic diagram of a provident fund new home loan workflow provided in one embodiment of the present application.

[0026] Figure 4 A schematic diagram of a single-round workflow provided in an embodiment of the present application.

[0027] Figure 5 A schematic diagram of a multi-round workflow provided in an embodiment of the present application.

[0028] Figure 6 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] The purpose of this application is to solve the challenges of large language model hallucination and unstable workflow calls in the prior art, and to complete workflow tasks efficiently and accurately through human-machine collaboration. This involves a variety of technologies such as artificial intelligence, human-computer interaction, workflow management, and intelligent agents. A multi-agent system is a system composed of multiple intelligent agents with autonomous decision-making capabilities. Workflow technology focuses on modeling, managing, and optimizing business processes. Human-machine collaboration aims to give full play to the advantages of humans and machines and achieve efficient cooperation.

[0031] This application uses both GUI and conversational user interface (CUI) to implement graph workflows, which have high freedom and good determinism. It can provide flexible questions and answers and accurate work. The review can be completed by automatic review agents, or confirmed by humans when necessary.

[0032] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0033] The workflow operation method based on multi-agent and human-machine collaboration provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the business scenarios and business requests to be processed to the server 104. After receiving the business scenarios and business requests to be processed, the server 104 splits the business scenarios into workflow tasks, builds a workflow model, and executes the workflow based on the workflow model, the human-computer collaborative multi-agent system and the business request. The server 104 can feedback the execution results of the workflow to the terminal 102. In addition, in some embodiments, the workflow operation method based on multi-agent and human-computer collaboration can also be implemented by the server 104 or the terminal 102 alone.

[0034] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0035] In an exemplary embodiment, Figure 2 As shown, a workflow operation method based on multi-agent and human-machine collaboration is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or can be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used for explanation, and the steps include the following steps 201 to 205.

[0036] Step 201: split the business scenario into workflow tasks and build a workflow model.

[0037] In an exemplary embodiment, business experts or product designers manually convert business scenarios and business processes (such as loan processing processes, sales order processes, etc.) into business process diagrams, and then sort out the business process diagrams to obtain a workflow model that conforms to the graph workflow definition, thereby converting business execution into a workflow operation that flows between intelligent body nodes.

[0038] The workflow model is a directed cyclic graph, including a global state, a start node, a work node, a work edge, a jump condition, a business application program interface and a constraint condition.

[0039] The workflow model is formally expressed as ;in, For the workflow model, is the global state (including the global termination state ), is the starting node, For the working node, For the working side, is the jump condition, Provides business application programming interfaces, including database operations and global status renew.

[0040] The starting node The number of working nodes is one. Including worker agent nodes , Review Agent Node and end node . Worker Agent Node , Review Agent Node and end node The number is one or more, i is the index of the working agent node.

[0041] The working edge Include global state edges 、Human-machine collaboration , Ordinary process edge And the audit side. The audit side includes the audit passing side and review rejection The global state edge For the global state Flow edges during update.

[0042] The jump condition Bind to the working side The jump condition includes an unconditional jump and business application program interface jump . Working side There is only one business application interface jump with the business application interface The human-machine collaborative edge Only redirected by the business application program interface Binding. Can be unconditionally jumped And the business application program interface jump Binding. Review passed and review rejection side Only the unconditional jump Binding.

[0043] The constraints include the following 6 constraints.

[0044] (1) The working node With the global state One to one correspondence.

[0045] (2) When the normal process Bind the business application program interface jump When the normal process The incoming edge node triggers the business application program interface jump After that, jump to the normal process side The outgoing edge node and update the global state .

[0046] (3) When the human-machine collaboration Bind the business application program interface jump When the human-machine collaboration The incoming edge node triggers the business application program interface jump After that, the user confirms or corrects the error before jumping to the human-machine collaboration edge. The outgoing edge node and update the global state .

[0047] (4) The unconditional jump Only one incoming edge node and one outgoing edge node are bound.

[0048] (5) The business application program interface jumps Can bind any incoming edge node and any outgoing edge node.

[0049] (6) The audit agent node There is only one in-degree unconditional edge, the audit agent node The out-degree edge of can only be a review edge.

[0050] All nodes in the workflow model are agents. They are broadly defined agents such as to-do assistants, database question-answering, and autonomous driving control. Agents are large language models that can call application programming interfaces. They can not only communicate, but also transmit information and data to the actual system.

[0051] Agent nodes (including start nodes, work agent nodes, review agent nodes and end nodes) can independently complete the tasks of the current workflow, such as automatic information collectors, automatic application submitters, automatic contract reviewers, etc. Specifically, agent nodes can understand user intentions and jump conditions, and thus determine through dialogue whether the user wants to jump (by determining whether the "large language model output contains the calling tool field"), where to jump (by determining the specific called tool to determine the jump target), and whether the conditions are met.

[0052] Step 202: Initialize a human-machine collaborative multi-agent system based on the workflow model. The human-machine collaborative multi-agent system includes multiple large language model agents. The multiple large language model agents serve as start nodes or working nodes in the workflow model.

[0053] Specifically, according to the workflow model , realize some of the agents, realize the application program interface and human-computer collaborative logic they rely on. Then start the human-computer collaborative multi-agent system.

[0054] In an exemplary embodiment, the human-machine collaborative multi-agent system can complete complex workflow tasks. The human-machine collaborative multi-agent system includes a dialogue human-machine collaboration layer, a dialogue application program interface call layer, and a dialogue process monitoring layer.

[0055] The dialogue human-machine collaboration layer is responsible for managing and storing the global and single-round dialogue status, controlling the workflow flow, controlling the API call, and controlling the human-machine collaborative operation. Workflow flow control is divided into single-round workflow flow control, multi-round workflow flow control, and updating the workflow status to the persistent database. API call control is divided into user business field identification, business field query, and business processing result notification. Human-machine collaborative operation control is divided into parameter confirmation and parameter modification before the API call, as well as the global review of the dialogue process monitoring layer.

[0056] The conversational API call layer includes workflow-related database operation APIs, business processing APIs, and other third-party APIs. The API service will connect to the database when necessary.

[0057] The dialogue process monitoring layer includes key information query and information review in the global process. When the review is passed, the global status is updated to the next global workflow node, and when the review is rejected, the global status is rolled back to the previous global workflow node.

[0058] In a specific example, Figure 3 As shown in the figure, in the provident fund new house loan business, the human-machine collaborative multi-agent system includes the following agents.

[0059] 1) Realize loan business The following working nodes are required 。

[0060] Ordinary agents: review query agent and contract query agent.

[0061] Human-machine collaborative agent: information confirmation agent, loan application agent, application cancellation agent, contract signing agent.

[0062] Review agent: application review agent, loan review agent.

[0063] Each agent can be bound to multiple tasks (connected to other working nodes through work edges). Users can chat with the agent. When the user wants to complete a task (jump to the next node), the agent determines whether the task conditions are met, that is, whether the work edge business application program interface jump is satisfied. The input parameters can be determined by general tool parameter tuning methods.

[0064] In this example, each agent is a chatbot with a task (work side) chatting with the user. During the conversation, the agent will remind the user to complete the task (work side flow, such as reminder to collect information, reminder to upload materials, etc.), and then judge whether the task conditions are met (whether the business application program interface jump of the work side is met) based on the conversation history. ), if the user wants to complete the task in the dialogue (such as using the general agent prompt word intention judgment technology), the workflow will flow to the next node when the task conditions are met.

[0065] 2) Realize loan business The following business application program interface jump is required .

[0066] Non-cooperative business application interface jump ： Application policy query API, application query API, application cancellation API, loan calculation API, contract signing query API, contract query API, this month's bill API, next month's bill API.

[0067] Collaborative business application program interface jump : Confirmation information API, application submission API, loan signing API.

[0068] 3) Realize loan business The steps of human-machine collaboration are as follows.

[0069] Loan information confirmation: In the dialogue, when the user confirms that he is ready to take out a loan, he first confirms special information such as marital status, retirement status, etc. in the loan conditions.

[0070] Submit application confirmation: In the dialogue, when the user decides to submit the application materials that have been uploaded, the system will pop up a window to ask the user to confirm the materials, and then submit the application materials after confirmation.

[0071] Application materials review: After the user submits the application, the reviewer reviews the application materials submitted by the user; during the review period, the user cannot proceed to the next process.

[0072] Loan information confirmation: In the conversation, when the user decides to sign a loan contract, the loan form is displayed to allow the user to confirm the loan amount, time, and repayment method, and then submit the loan contract information after confirmation.

[0073] Contract signing and loan approval: After the user signs the contract, the auditor will review the contract information submitted by the user and approve the loan if it is correct. During the review period, the user cannot proceed to the next process.

[0074] Step 203, determine the workflow type according to the service request input by the user, and determine the starting agent node according to the workflow type. The workflow type is a single-round workflow or a multi-round workflow. The starting agent node is a starting node or a working agent node.

[0075] Specifically, if the workflow type is a single-round workflow, the start agent node is the starting node If the workflow type is a multi-round workflow, the user's current global state is obtained from the persistent database. , and use the corresponding working agent node as the starting agent node .

[0076] like Figure 4 As shown, in a single-round workflow, starting from the start node, multiple working agent nodes and application program interfaces perform human-machine collaborative or non-collaborative operations, and finally end the process at the end node. Among them, the working agent node determines the next working agent node based on the jump condition.

[0077] like Figure 5As shown, in a multi-round workflow, starting from the start node, multiple working agent nodes are determined according to the global state, and the multiple working agent nodes then perform human-computer collaborative or non-collaborative operations with the application interface respectively, among which some working agent nodes can interact with the review agent node, and finally the process ends at the end node.

[0078] In a specific example, in the multi-round workflow of a provident fund new home loan: when a new user enters the conversation, the user starts from the "not applied" state; old users need global state management to first determine the current global state and correspond to the working agent node, and then initialize from the "applying", "signing", "loaning" and other states.

[0079] Step 204, based on the workflow model and the human-machine collaborative multi-agent system, starting from the start agent node, the workflow is executed between each working node.

[0080] Specifically, according to the workflow model , starting from the agent node Start to transfer tasks between various working agent nodes. For each working agent node, the human-machine collaborative multi-agent system will jump according to the bound jump conditions. , call the corresponding business application program interface to perform business processing.

[0081] In a specific example, when in the "unconfirmed" state, the user can ask questions about the application conditions. Since the "Application Policy Query API" is a non-collaborative API, the information confirmation agent will directly (rather than human-computer collaboration) call the API and answer based on the call results.

[0082] When in the "applying" state, the user can cancel the loan application. Since the "Cancel Application API" is a non-collaborative API, the application query agent will directly call this API and answer based on the call result.

[0083] That is to say, when the flow is transferred to a specific agent in the human-machine collaborative multi-agent system, the agent calls , when it is a non-collaborative application interface, the corresponding application interface is directly called to answer, and this agent can be called a normal agent.

[0084] Step 205, during the workflow execution process, human-machine collaboration is performed based on the human-machine collaboration edge, business audit is performed based on the audit edge and the audit agent node, and the global state is updated until the global state is the global termination state to end the process.

[0085] Furthermore, step 205 specifically includes steps 301 to 303 .

[0086] Step 301: During the workflow execution process, if a human-machine collaborative edge is encountered, wait for the user to make a second confirmation, update the global state according to the user confirmation result, and determine the next execution direction according to the jump condition.

[0087] Specifically, during the workflow execution, if there is a human-machine collaborative edge that requires human-machine collaborative processing, , the human-machine collaborative multi-agent system will call up the human-machine collaborative module (such as pop-up windows, forms, physical buttons, etc.), and continue the execution process after the user's second confirmation. The human-machine collaborative multi-agent system updates the global state according to the returned results and determines the next execution direction according to the jump conditions.

[0088] In a specific example, when in the "not applied" state, the user can confirm the application and submit the application materials. At this time, the application agent will pop up a window to prompt the user to confirm (instead of submitting directly) and call the corresponding "Submit Application Materials Application Programming Interface" to complete the application.

[0089] When in the "Signing" state, the user can propose to sign the contract. At this time, the contract signing intelligent body will pop up a signing information form for the user to edit it again. After the user edits and confirms, the corresponding "Loan Signing Application Programming Interface" is called to complete the signing.

[0090] When the agent to which the flow is transferred needs to call When it is a collaborative application interface, the user is asked to edit and confirm the operation again, and then call and answer it after the operation is completed. This agent can be called a human-computer collaborative agent.

[0091] Step 302, during the workflow execution process, if an audit edge is encountered or the audit agent node is executed, wait for the auditor to conduct a business audit, and update the global state according to the audit result, and advance the workflow to the next work node or roll back to the previous work node.

[0092] Specifically, when the workflow executes to the audit agent node When the user submits an application, the human-machine collaborative multi-agent system can provide the auditor with the information submitted by the user. According to the business situation, the auditor can choose to approve or reject the application. The human-machine collaborative multi-agent system updates the global status based on the audit results and advances the workflow to the next work node or rolls back to the previous work node.

[0093] Step 303, when the workflow is executed to the end node, if the workflow type is a single-round workflow, the global state is updated to the global termination state to end the process. If the workflow type is a multi-round workflow, the global state is updated and continues to execute from the start node until the global state is the global termination state to end the entire process.

[0094] Specifically, when the workflow executes to the end node If it is a single-round workflow, the workflow ends; if it is a multi-round workflow, the global state is updated. , and continue from the starting node Execute until the global termination state , then the whole process ends.

[0095] Furthermore, the workflow operation method based on multi-agent and human-machine collaboration also includes step 206.

[0096] Step 206, continuously monitor the human-machine collaborative multi-agent system, and collect and analyze workflow execution data, application interface call data and user behavior data to evaluate the performance of the human-machine collaborative multi-agent system.

[0097] Specifically, the human-machine collaborative multi-agent system is continuously monitored, and the workflow execution data, application interface call data, user behavior data, etc. are collected and analyzed to evaluate the performance and effect of the human-machine collaborative multi-agent system. If the workflow encounters errors or abnormal situations during execution, the human-machine collaborative multi-agent system will also handle the exceptions and record the error information for subsequent analysis and improvement.

[0098] This application focuses on the business scenario-based collaboration model between humans and intelligent agents, aiming to achieve a more efficient workflow: 1) Artificial intelligence agents are designed to have the ability to execute complex workflows to ensure that they can perform tasks in various scenarios. 2) In the interaction between humans and intelligent agents, they use a "hand-in-hand" approach, that is, combining GUI and CUI, rather than relying solely on verbal instructions, to implement precise and effective control of the intelligent agent. This interaction method ensures the accuracy and efficiency of the collaboration between humans and intelligent agents.

[0099] This application optimizes the multi-agent system: 1) supports graph workflows; 2) supports human-machine collaboration. The graph workflow defined in this application expresses the agent (node), control operation (edge), and business process (graph layout) in the same work graph. The workflow can be completed by following the flow of node and edge definitions. It can run complex workflows and achieve human-machine collaboration. The workflow-running robot designed in accordance with this application has the ability to complete tasks autonomously and can be precisely controlled. It can easily achieve complex workflows such as "customer acquisition, conversion, and order placement": customer acquisition, chatting to obtain leads; conversion, sales business process; order placement, user confirmation of payment. The implementation method and function comparison of the workflow-running robot designed in this application and common robots are shown in Table 1.

[0100] Table 1 Robot implementation and function comparison

[0101]

[0102] This application significantly improves the efficiency and accuracy of the human-machine collaborative multi-agent system and optimizes the user experience through the following technical points.

[0103] 1. Human-machine collaborative information verification: When the multi-agent system decides to call the tool to submit key data, it "interrupts the model casting" and transfers the verification right to the user, and can also adjust the input parameters of the tool for a second time. By introducing the user verification link, the accuracy and reliability of the multi-agent system's modification of key information is ensured, effectively avoiding the "hallucination" phenomenon caused by the large language model, ensuring the accuracy and reliability of key data, and directly improving the accuracy of the human-machine collaborative multi-agent system's decision-making, thereby improving the reliability of the overall workflow.

[0104] 2. Human-machine collaborative single-round flow management: When the user interacts with the multi-agent system only once, a directed cyclic graph describing the single-agent collaboration is constructed to describe the single-agent collaboration. Subgraphs can also be nested, so that the human-machine collaborative multi-agent system can clearly perform complex tasks such as nesting and looping in the directed cyclic graph. This not only improves the flexibility and adaptability of the human-machine collaborative multi-agent system, but also greatly simplifies the management and execution of complex workflows and improves overall work efficiency.

[0105] 3. Global state management of human-machine collaboration: When users interact with the multi-agent system multiple times, a global multi-round workflow state management with persistent storage is built. The human-machine collaborative multi-agent system can coordinate the flow and operation between agents in multiple rounds of interaction. The global state and single-round flow work together to complete the workflow tasks. Users / auditors can manage the global multi-round workflow state, ensuring that the agents in the workflow automatically complete the tasks within a controllable range, solving the problem of global control difficulties in the existing technology, and enabling the human-machine collaborative multi-agent system to better adapt to complex workflow scenarios.

[0106] 4. Human-machine collaborative interaction mode: Through the combination of active verification, directed cyclic graphs and global management, the "multi-agent-user-multi-agent" human-machine collaborative interaction mode is realized. This interaction mode not only improves the usability and user experience of the human-machine collaborative multi-agent system, but also lowers the technical threshold, so that users who lack relevant background knowledge can also operate easily.

[0107] In summary, this application significantly improves the efficiency, stability and accuracy of workflow operation and optimizes user experience by introducing core strategies such as information verification, single-round flow management, global state management and interactive mode of human-machine collaboration. Therefore, this application has broad application prospects and huge market potential in business processing, customer service, intelligent manufacturing and other fields.

[0108] This application integrates multi-agent system and human-machine collaboration into the workflow model, completes complex workflow tasks, optimizes user business experience, and improves the efficiency of the overall workflow. Through strategies such as information verification, single-round flow management, global state management, and human-machine collaborative interaction mode, a workflow design and operation system based on multi-agent and human-machine collaboration can be constructed.

[0109] This application successfully breaks through the limitation of thinking that large language model applications are mostly in the form of chatbots, and puts forward the following innovative ideas: 1) Multi-agent systems should have the ability to execute complex workflows; 2) Multi-agent systems should provide new ways of interaction: combining graphical user interfaces and conversational user interaction interfaces. It not only enriches the application scenarios of artificial intelligence agents, but also improves the efficiency and accuracy of interaction between humans and agents. It solves the problems of unreliable content generation, single multi-agent cooperation mode, difficult global workflow control, and high technical barriers.

[0110] The present application also provides an application scenario, which applies the above-mentioned workflow operation method based on multi-agent and human-machine collaboration. Specifically: The workflow operation method based on multi-agent and human-machine collaboration provided in this embodiment can be applied in the sales order scenario. In the sales order scenario, first, a workflow model is constructed according to the business process of the sales order scenario, and then based on the workflow model, a human-machine collaborative multi-agent system is used to execute the sales order business flow, and the corresponding execution results are obtained through human-machine collaboration.

[0111] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 6As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store business scenarios. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a workflow operation method based on multi-agent and human-computer collaboration is implemented.

[0112] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0113] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0114] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0115] 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, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0116] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0117] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0118] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0119] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A workflow operation method based on multi-agent and human-machine collaboration, characterized in that: The workflow operation method based on multi-agent and human-machine collaboration includes: The business scenario is divided into workflow tasks, and a workflow model is constructed; the workflow model is a directed cyclic graph, including a global state, a start node, a work node, a work edge, a jump condition, a business application program interface, and a constraint condition; wherein the work node includes a work agent node, an audit agent node, and an end node; the work edge includes a global state edge, a human-machine collaboration edge, a common process edge, and an audit edge; the jump condition includes an unconditional jump and a business application program interface jump; Based on the workflow model, a human-machine collaborative multi-agent system is initialized; the human-machine collaborative multi-agent system includes a plurality of large language model agents; the plurality of large language model agents serve as start nodes or working nodes in the workflow model; The human-machine collaborative multi-agent system includes a dialogue human-machine collaboration layer, a dialogue application program interface call layer, and a dialogue process monitoring layer; The dialogue human-machine collaboration layer is responsible for managing and storing the global and single-round dialogue status, controlling the workflow flow, controlling the API call, and controlling the human-machine collaborative operation; the workflow flow control is divided into single-round workflow flow control, multi-round workflow flow control, and updating the workflow status to the persistent database; the API call control is divided into user business field identification, business field query, and business processing result notification; the human-machine collaborative operation control is divided into parameter confirmation and parameter modification before the API call, and the global review of the dialogue process monitoring layer; The dialogue API call layer includes workflow-related database operation APIs, business processing APIs, and other third-party APIs; The dialogue process monitoring layer includes key information query and information review in the global process; when the review is passed, the global status is updated to the next global workflow node, and when the review is rejected, the global status is rolled back to the previous global workflow node; Determine the workflow type according to the business request input by the user, and determine the starting agent node according to the workflow type; the workflow type is a single-round workflow or a multi-round workflow; the starting agent node is a starting node or a working agent node; Based on the workflow model and the human-machine collaborative multi-agent system, starting from the start agent node, the workflow is transferred and executed between each working node; During the workflow execution process, human-machine collaboration is performed based on the human-machine collaboration edge, business audit is performed based on the audit edge and the audit agent node, and the global state is updated until the process ends when the global state is a global termination state.

2. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: The number of the starting node is one; the number of the working agent nodes, the auditing agent nodes and the ending nodes is one or more.

3. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: The jump condition is bound to the working edge; the global state edge is the flow edge when the global state is updated; the human-machine collaboration edge is only bound by the business application program interface jump; the ordinary process edge can be bound by the unconditional jump and the business application program interface jump; The audit edge is only bound by the unconditional jump.

4. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: The audit edges include audit-passed edges and audit-rejected edges.

5. The workflow operation method based on multi-agent and human-machine collaboration according to claim 3 is characterized in that: The constraints include: The working nodes correspond to the global state one by one; When the common flow edge is bound to the business application program interface jump, after the business application program interface jump is triggered by the input edge node of the common flow edge, the common flow edge is jumped to the output edge node of the common flow edge, and the global state is updated at the same time; When the human-machine collaborative edge is bound to the business application program interface jump, after the input edge node of the human-machine collaborative edge triggers the business application program interface jump, user confirmation or correction is performed first, and then the output edge node of the human-machine collaborative edge is jumped, and the global state is updated at the same time; The unconditional jump is only bound to one incoming edge node and one outgoing edge node; The business application program interface jump can bind any incoming edge node and any outgoing edge node; The audit agent node has a unique in-degree unconditional edge, and the out-degree edge of the audit agent node is an audit edge.

6. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: If the workflow type is a single-round workflow, the starting agent node is the starting node; If the workflow type is a multi-round workflow, the current global state of the user is obtained from the persistent database, and the corresponding working agent node is used as the starting agent node.

7. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: During the workflow execution process, human-machine collaboration is performed based on the human-machine collaboration edge, business audit is performed based on the audit edge and the audit agent node, and the global state is updated until the global state is the global termination state, and the process is terminated, specifically including: During the workflow execution process, if a human-machine collaborative edge is encountered, the user is waited for secondary confirmation, and the global state is updated according to the user confirmation result, and the next execution direction is determined according to the jump condition; During the workflow execution process, if an audit edge is encountered or the audit agent node is executed, the audit agent waits for the auditor to conduct a business audit, and updates the global state according to the audit result, and advances the workflow to the next work node or rolls back to the previous work node; When the workflow executes to the end node, if the workflow type is a single-round workflow, the global state is updated to the global termination state to end the process; if the workflow type is a multi-round workflow, the global state is updated and continues to execute from the start node until the global state is the global termination state to end the entire process.

8. The workflow operation method based on multi-agent and human-machine collaboration according to claim 1 is characterized in that: The workflow operation method based on multi-agent and human-machine collaboration also includes: The human-machine collaborative multi-agent system is continuously monitored, and workflow execution data, application program interface call data, and user behavior data are collected and analyzed to evaluate the performance of the human-machine collaborative multi-agent system.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the workflow operation method based on multi-agent and human-computer collaboration as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the workflow operation method based on multi-agent and human-computer collaboration described in any one of claims 1 to 8 is implemented.

Images