Intelligent processing method and system for business process decision nodes based on AI Agent

Abstract

The present invention belongs to the technical field of business process management, and specifically relates to an intelligent processing method and system for business process decision nodes based on AI Agent. The method includes: S1, selecting a large model base according to actual needs; S2, transforming the rule engine of the process, and combining the intelligent body capabilities of the large model base to complete the construction of the rule robot intelligent body; S3, setting clear exit conditions for the process rule robot intelligent body. Before the exit conditions are met, the process rule robot intelligent body will cyclically call the tool set, and use the current execution result as supplementary data for the next reflection, and finally complete all agreed transactions. The present invention has the characteristics of being able to significantly improve the flexibility, automation rate and decision-making accuracy of business processes.

Description

Technical Field

[0001] The present invention belongs to the technical field of business process management, and specifically relates to an AI Agent-based business process decision node intelligent processing method and system. Background Art

[0002] In traditional business process management and automation systems, when a process model faces multiple branch path choices, conditional decision nodes are usually pre-configured.

[0003] Conditional decision nodes are described as "Gateways" in the BPMN2.0 specification and can be further subdivided into different types such as exclusive gateways, parallel gateways, and compatible gateways.

[0004] These decision rules can be based on simple conditional judgments or involve complex logical operations. The more common technical implementation methods are: 1. Direct rule logic based on conditional expressions 2. Decision table / decision matrix 3. Decision Tree 4. Integrate external rule engines (such as Drools, ILOG JRules, etc.) 5. Customize complex scripts or services Although the complexity of the above technical solutions varies, in essence, they all require delivery engineers to conduct a lot of demand research during the process modeling phase, sort out and abstract appropriate business rules, and preset them into the process model. The subsequent process instance execution process will strictly follow these preset rules.

[0005] The scheme of pre-conditional decision nodes has the following two main disadvantages: 1. The investigation and sorting out of preset rule logic requires a lot of time and manpower costs, and may even become the longest work item in the delivery of process management projects.

[0006] 2. Since all rules need to be pre-configured in the process model, the subsequent process execution process needs to strictly abide by these rule logics. If the actual business changes and the rules need to be adjusted accordingly, the only way is to modify the process model and re-release the version. For some businesses with frequent iterations, this model cannot meet their requirements for flexibility.

[0007] For the selection of another part of multi-branch paths, after process investigation, it is considered that automatic judgment cannot be made through preset rules, so it is necessary to introduce a manual processing link, and the subsequent process direction needs to be manually determined by the processing personnel.

[0008] The human intervention approach has two main disadvantages: 1. Additional manual processing links are added to the process chain, reducing the automation rate of the process. If a busy leader is the person in charge of the link, it may even form a "process black hole" and greatly reduce the efficiency of the process operation.

[0009] 2. The process is manually determined by branch paths, and its efficiency and accuracy often depend on the business experience and decision-making ability of the person in charge. This introduces uncertainty into the operation of the process and increases the risk of process errors.

[0010] Traditional business process management platforms have many shortcomings when dealing with multi-branch path selection. With the vigorous development of generative AI technology, this application attempts to solve these problems with the semantic understanding ability of large models and AI Agent technology.

[0011] Therefore, it is very important to design an AI Agent-based intelligent processing method and system for business process decision nodes that can significantly improve the flexibility, automation rate and decision-making accuracy of business processes. Summary of the invention

[0012] The present invention aims to overcome the problems in the prior art of traditional process rule engines, such as heavy preliminary research, inflexible model presets, low efficiency of human intervention and high error rate of manual decision-making, and provides an AI Agent-based business process decision node intelligent processing method and system that can significantly improve the flexibility, automation rate and decision-making accuracy of business processes.

[0013] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions: The intelligent processing method of business process decision nodes based on AI Agent includes the following steps; S1, select a large model base according to actual needs; S2, transform the process rule engine and combine it with the intelligent agent capabilities of the large model base to complete the construction of the process rule robot intelligent agent; S3, sets clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

[0014] Preferably, in step S1, the large model base has general semantic understanding capabilities and Agent intelligent body construction capabilities; the Agent intelligent body construction capability refers to making task scheduling decisions through autonomous thinking by analyzing existing information.

[0015] Preferably, step S1 comprises the following steps: S11, based on the open source model ChatGLM3-6B-base, uses the process data tags accumulated in several business process platform delivery projects to perform LoRA fine-tuning on the open source model, and re-quantifies and publishes the open source model into the "AlphaFlow large process model" suitable for private deployment; the process data tags are desensitized.

[0016] Preferably, step S2 comprises the following steps: S21, determining the process data needed as a basis for decision making; the process data includes process form content, process instance information, and input and output parameter values ​​of the process robot; S22, define the tool tool in the tool use mode. The specific process is as follows: In the Agent, all possible branch paths of the process are abstracted as tools. Considering that each branch path must lead to a specific target node, the corresponding tool feature definition is given by analyzing the role of the corresponding target node to prepare for the decision-making judgment of the Agent. The target node includes the manual processing link, the process robot link, the sub-process link and the end link. S23, in the Agent intelligent body mode, inherits the rules and conditions that have been preset in the process of building the large process model, and is used as the data basis for the Agent intelligent body to plan; S24, integrates process data, tool definitions and preset rule condition contents, combines to build a complete prompt word project, and drives the operation of the large process model.

[0017] Preferably, in step S21, the process instance information includes link handler information, link authority information and link decision.

[0018] Preferably, in step S22, the target node includes a manual processing link, a process robot link, a sub-process link and an end link.

[0019] Preferably, step S3 comprises the following steps: S31, the process rule robot agent makes a decision based on the existing data and drives the process engine to call one or more process robots to perform tasks; S32, the process rule robot agent observes and verifies the validity of the task completion result. If the result does not meet expectations, a new strategy is formulated based on the current situation and task scheduling is initiated again; S33, repeating the process from step S31 to step S32 until it is confirmed that the task is accomplished.

[0020] The present invention also provides an AI Agent-based business process decision node intelligent processing system including: Selection module, used to select the large model base according to actual needs; The intelligent agent construction module is used to transform the process rule engine and combine the intelligent agent capabilities of the large model base to complete the construction of the process rule robot intelligent agent; The execution module is used to set clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) Reducing the time cost of preliminary investigation and configuration: The present invention can learn, optimize and adjust rules during the process operation through the "planning" and "reflection" modes of AI Agent, reducing a large amount of preliminary investigation and configuration work and significantly shortening the project delivery cycle; (2) Improving the flexibility of business processes: The present invention can dynamically evaluate and learn during the process operation through AI Agent technology, make optimal decisions based on real-time conditions, and greatly improve the flexibility and adaptability of business processes; (3) Improving the automation rate of processes: The present invention can realize the automatic selection, scheduling and execution of subsequent paths of the process through the "tool use" mode of AI Agent, greatly improving the automation rate of business processes; (4) Improving the accuracy and efficiency of decisions: The present invention uses the semantic understanding of large models and the decision-making and scheduling capabilities of AI Agent to more accurately analyze process data and make more scientific and reasonable decisions; compared with traditional manual intervention, the decision-making process of AI Agent is not only faster but also more accurate, effectively reducing the risk of process errors; (5) The present invention uses AI Agent to automatically select, schedule and execute subsequent paths of the process. The integration of Agent technology into the business process rule engine not only solves many defects in traditional process management, but also significantly improves the flexibility, automation rate and decision-making accuracy of business processes, providing strong technical support for the digital transformation of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A principle block diagram for constructing a rule-based robot agent in the present invention; Figure 2 A schematic diagram of the decision-making process of the Agent in the present invention; Figure 3 It is a schematic diagram of a closed-loop application of the Agent intelligent body decision-making process in the present invention in an actual process scenario; Figure 4 This is a business process management diagram under the traditional process engine mode; Figure 5 The present invention is a schematic diagram of business process management using the intelligent processing method of business process decision nodes based on AI Agent. DETAILED DESCRIPTION

[0023] In order to more clearly illustrate the embodiments of the present invention, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings and other implementation methods can be obtained based on these accompanying drawings without creative work.

[0024] The present invention provides an intelligent processing method for business process decision nodes based on AI Agent, comprising the following steps: 1. Choose a large model base according to actual needs; 2. Transform the rule engine of the process and combine it with the intelligent agent capabilities of the large model base to complete the construction of the rule process robot intelligent agent; 3. Set clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

[0025] For step 1, the specific process is as follows: The large model base is the cornerstone of the process engine agent transformation and directly affects the overall performance of the engine. The basic large model should have the following characteristics: (1) Possess general semantic understanding capabilities (necessary feature) The basic large model should have strong general semantic understanding capabilities. This is usually related to factors such as the training data quality, model architecture, and parameter scale of the large model, and should be carefully considered in combination with the budget cost.

[0026] (2) Possess AI Agent capabilities (necessary characteristics) The basic large model must have strong agent intelligent body construction capabilities, that is, it can analyze existing information and make decisions on task scheduling by thinking independently. Some models that only have natural language chat dialogue capabilities are not applicable.

[0027] (3) Targeted training and learning of process domain data (bonus point) Since the large model base is deeply embedded in the process engine and interacts with the business data in the process field, if the model is pre-trained and fine-tuned with process field data, its performance in actual business scenarios will be greatly improved.

[0028] For example, process data from hundreds of business process platform delivery projects (including form content, instance information, job title information of the parties handling the links, etc., all of which have been desensitized to ensure privacy and security) are collected and used as training data sets.

[0029] Next, LoRA (Low-Rank Adaptation) is selected as the fine-tuning solution. The LoRA solution can reduce the number of parameter updates by introducing low-rank matrix decomposition in the pre-trained model, thereby reducing computing resource consumption and training time. Practice has proved that LoRA is an efficient fine-tuning solution, which enables the pre-trained model to quickly learn the key features of the process data with relatively low resource consumption.

[0030] (4) Other personalized private knowledge (bonus points) When this technical solution is delivered to the business process platform of a specific enterprise, that is, when the model is applied to the business scenarios of a specific enterprise, if the model can be pre-trained and fine-tuned with the personalized private knowledge of the enterprise, or if the enterprise's internal knowledge base data can be called in real time through RAG, KG and other technologies during the operation of the model, it will significantly enhance the business understanding ability of the process intelligent body.

[0031] For example, real-time calling of enterprise internal knowledge base data through RAG (Retrieval-Augmented Generation) technology during model operation is the mainstream solution provided in process products. The Milvus vector database is introduced as a carrier for storing and managing text vectors in the knowledge base. By converting the text in the knowledge base into a vector representation, semantic-based similarity matching (rather than just traditional keyword matching) can be achieved.

[0032] During the operation of the process agent, the big model can dynamically retrieve relevant knowledge items from the enterprise's internal knowledge base and incorporate them into the generated results. For example, when processing process decisions, the agent can retrieve relevant enterprise rules and regulations, historical cases and other information from the knowledge base based on the current process data, so as to make more accurate decisions.

[0033] Taking into account the above four characteristics and the budget investment of enterprises of different sizes in system construction, the present invention provides customers with a variety of large model base options in the product solution.

[0034] At present, the most widely used solution of this invention is: based on the open source model ChatGLM3-6B-base, the process data tags (which have been desensitized) accumulated in hundreds of business process platform delivery projects are used to fine-tune the LoRA, and then re-quantified and published as the "AlphaFlow Large Process Model (AlphaFlow LPM)" suitable for private deployment. This serves as the large model base of this technical solution.

[0035] The selection of the large model base is the basis of the entire technical solution. It provides core computing power and intelligent support for the subsequent construction of process rule robots and the operation of process intelligent bodies.

[0036] For step 2, the specific process is as follows: After determining the large model base, the present invention begins to transform the rule engine of the process and build a process rule robot with AI Agent capabilities.

[0037] The construction process generally has the following key points and steps: (1) Determine the process data needed as a basis for decision making Sort out and determine which process data may have an impact on the Agent's decision-making. The data usually includes: process form content, process instance information (stage handler information, stage authority information, stage decision, etc.), process robot input and output parameter values, etc. These data will become an important basis for the Agent to plan (Planning) and reflect (Reflection).

[0038] (2) Give the conditions that trigger the corresponding branch paths one by one This step is essentially the process of defining the tool in the tool use mode. In the Agent, the possible branch paths of the process are abstracted as tools. Since each path must lead to a specific node (which may be a manual processing link, a process robot link, a sub-process link, an end link, etc.), the feature definition of the tool can be given by analyzing the role of the target node to prepare for the Agent's decision-making.

[0039] (3) Combined with the preset rule logic in the process model In the process modeling stage, some very clear rule conditions may have been set up through traditional methods (such as preset conditional expressions, etc.), so these logics need to be inherited in the Agent mode and used as the data basis for the Agent to plan. This can maximize compatibility with the paradigm of traditional rule engines.

[0040] (4) Integrate the above three parts to build a complete prompt word project to drive the operation of the large process model.

[0041] The process rule robot is constructed by integrating the three parts of information, namely process data, tool definition, and preset rules, and combining the agent capabilities of the large process model. The agent can independently think and judge which branch paths should be selected based on this information, and automatically complete the entire process of process driving.

[0042] like Figure 1 As shown, the results of the two key points of step 1 and 2 and their relationship are intuitively described. The large model base selected by the present invention is Figure 1 The "AlphaFlow large process model" in the paper is a new type of process model. On this basis, the present invention comprehensively applies process data, path (tool) definition, preset rules and other information to construct a prompt word project to drive the operation of the large process model. The two together construct a rule-based robot agent.

[0043] For step 3, the specific process is as follows: After completing the technical implementation of the key points in the above two steps, the process agent is ready for actual operation.

[0044] For most business scenarios, the Agent can meet the requirements of driving the process by making a path decision once. However, in some complex scenarios, especially in the automation scenarios combined with process robots, a single flow decision may not be able to solve all problems. At this time, it is necessary to set clear exit conditions for the agent. Before the conditions are met, the agent will call the tool set cyclically and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

[0045] This mechanism effectively simulates the entire process of humans achieving their goals through continuous thinking, iteration, and gradual evolution, and also embodies the technical essence of AI Agent.

[0046] Figure 2 It is a streamlined Agent decision-making process: like Figure 2 As shown in the figure, the agent makes an analysis plan based on the existing data information, determines the strategy, and then selects the appropriate tool to take action. During the action process, the agent observes the execution results, judges the task completion rate, and constantly adjusts the plan, makes new decisions, and starts new actions. This cycle repeats, constantly approaching the optimal solution until the task goal is achieved.

[0047] The above-mentioned intelligent agent decision-making process idea is combined with the process scenario, which can be as follows Figure 3 The closed-loop application of the thought chain with the effect shown: like Figure 3As shown in the figure, after the process starts, the AI ​​Agent rules agent makes a plan based on the existing data, and then drives the process to take action, that is, calling one or more process robots to execute. Then, the agent perceives the execution results, observes and verifies the effectiveness of the task completion results. If there is still a gap from the expectation, a new strategy is formulated based on the current situation, and task scheduling is initiated again (i.e. Figure 3 Repeat this process until the task is confirmed to be completed (i.e. Figure 3 "The result is valid and the task is completed" in the task).

[0048] In addition, the present invention also provides an AI Agent-based business process decision node intelligent processing system including: Selection module, used to select the large model base according to actual needs; The intelligent agent building module is used to transform the rule engine of the process and combine the intelligent agent capabilities of the large model base to complete the construction of the rule robot intelligent agent; The execution module is used to set clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

[0049] In order to more clearly describe the component generation process, the present invention gives a simple example to further illustrate and practice the above technical solution: When a pre-sales engineer needs assistance from the technical team at work, he or she usually applies through the "Pre-sales Project Technical Support Process" and describes the problem encountered in the process. The next step of the process will be approval by the company's technical director, who will determine the type of problem (secondary development, solution design, operation and maintenance deployment, etc.) based on the applicant's needs description and guide the process to the robot or sub-process that solves the corresponding problem.

[0050] Figure 4 This is the flowchart under the traditional process engine mode.

[0051] Depend on Figure 4 It can be seen that the assistance that pre-sales engineers may seek is diverse, and the most core task assignment of the process needs to be manually judged by the company's chief technical officer (CTO), which greatly reduces the execution efficiency of the process.

[0052] Next, the method of the present invention is used to transform the Agent of the rule engine of the above process.

[0053] The present invention introduces a "transaction type rule robot" into the process and enhances the rule engine in combination with AI Agent.

[0054] Figure 5 This is the flowchart after the Agent is introduced to perform AI enhancement on the process rule engine.

[0055] like Figure 5 As shown in the figure, after the applicant fills in the problem description, the rule robot can analyze the semantics of the demand text, determine the problem type, and automatically schedule the process path to one or more sub-processes (work contact form, operation and maintenance work form, product deployment application process, server solution application form, product demand processing process, technical task process and defect improvement application process) to solve the corresponding problem. This is a standard Planning + Tool use Agent process.

[0056] After such improvements, the process eliminates the human decision-making link, and the entire link no longer requires any human intervention, and has transformed into a truly fully automated process.

[0057] The innovative features of the present invention are as follows: 1. Selection, training and fine-tuning of the basic large model The comprehensive capabilities of the agent are directly proportional to the robustness of the four features of the basic model mentioned above. In addition to the selection of the basic model, it is also important to train, fine-tune, quantify, and republish the model using process data, which directly determines whether the basic model can be seamlessly integrated with the process engine.

[0058] 2. Iteration and fine-tuning of the relevant prompt word engineering that drives AI Agent Combining process data, tool definition, and preset rules, we build a prompt word project. And we need to combine it with the large model base, conduct continuous testing, iteration, and fine-tuning (with the help of tools such as LangFuse), and continuously improve its accuracy and hit rate in practice.

[0059] In addition, the same set of prompt word projects will perform differently under different large model bases, so they need to be maintained separately.

[0060] This process means a huge amount of technological accumulation barriers.

[0061] 3. Specific methods for integrating process rule engine and Agent technology Deeply integrating AI Agent technology with the process rule engine is the key to realizing intelligent business processes.

[0062] First, in the underlying architecture design, the process rule engine and AI Agent are decoupled, ensuring less intrusion into the process engine and higher compatibility with traditional models.

[0063] Secondly, design data interfaces and decision interfaces to ensure that process data can be smoothly transmitted to the AI ​​Agent, and that the Agent can efficiently feed back decision results to the rule engine.

[0064] Again, through data preprocessing and feature extraction, high-quality decision-making basis is provided for AI Agent.

[0065] Finally, through planning, reflection, tool use and multi-agent collaboration, dynamic evaluation and learning are achieved to make the best decision based on real-time situation.

[0066] This process requires not only sophisticated system design, but also continuous testing and optimization to ensure the stability and efficiency of the system, which means deep engineering barriers.

[0067] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent processing method for business process decision nodes based on AI Agent, characterized in that: The method comprises the following steps: S1, select a large model base according to actual needs; S2, transform the process rule engine and combine it with the intelligent agent capabilities of the large model base to complete the construction of the process rule robot intelligent agent; S3, sets clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

2. The intelligent processing method for business process decision nodes based on AI Agent according to claim 1 is characterized in that: In step S1, the large model base has general semantic understanding capabilities and Agent intelligent body construction capabilities; the Agent intelligent body construction capability refers to making task scheduling decisions through autonomous thinking by analyzing existing information.

3. The intelligent processing method for business process decision nodes based on AI Agent according to claim 2 is characterized in that: Step S1 includes the following steps: S11, based on the open source model ChatGLM3-6B-base, the open source model is fine-tuned with LoRA using process data tags accumulated in several business process platform delivery projects, and the open source model is re-quantified and released as the "AlphaFlow large process model" suitable for private deployment; The process data tags are desensitized.

4. The intelligent processing method for business process decision nodes based on AI Agent according to claim 3 is characterized in that: Step S2 includes the following steps: S21, determining the process data needed as a basis for decision making; the process data includes process form content, process instance information, and input and output parameter values ​​of the process robot; S22, define the tool tool in the tool use mode. The specific process is as follows: In the Agent, all possible branch paths of the process are abstracted as tools. Considering that each branch path must lead to a specific target node, the corresponding tool feature definition is given by analyzing the role of the corresponding target node to prepare for the decision-making of the Agent. The target node includes the manual processing link, the process robot link, the sub-process link and the end link. S23, in the Agent intelligent body mode, inherits the rules and conditions that have been preset in the process of building the large process model, and is used as the data basis for the Agent intelligent body to plan; S24, integrates process data, tool definitions and preset rule condition contents, combines to build a complete prompt word project, and drives the operation of the large process model.

5. The intelligent processing method for business process decision nodes based on AI Agent according to claim 4 is characterized in that: In step S21, the process instance information includes link handler information, link authority information and link decision.

6. The intelligent processing method for business process decision nodes based on AI Agent according to claim 4 is characterized in that: In step S22, the target node includes a manual processing link, a process robot link, a sub-process link and an end link.

7. The intelligent processing method for business process decision nodes based on AI Agent according to claim 4 is characterized in that: Step S3 includes the following steps: S31, the process rule robot agent makes a decision based on the existing data and drives the process engine to call one or more process robots to perform tasks; S32, the process rule robot agent observes and verifies the validity of the task completion result. If the result does not meet expectations, a new strategy is formulated based on the current situation and task scheduling is initiated again; S33, repeating the process from step S31 to step S32 until it is confirmed that the task is accomplished.

8. An AI Agent-based intelligent processing system for business process decision nodes, used to implement the AI ​​Agent-based intelligent processing method for business process decision nodes according to any one of claims 1 to 7, characterized in that: The AIAgent-based business process decision node intelligent processing system includes: Selection module, used to select the large model base according to actual needs; The intelligent agent construction module is used to transform the process rule engine and combine the intelligent agent capabilities of the large model base to complete the construction of the process rule robot intelligent agent; The execution module is used to set clear exit conditions for the process rule robot agent. Before the exit conditions are met, the process rule robot agent will cyclically call the tool set and use the current execution results as supplementary data for the next reflection, and finally complete all the agreed transactions.

Images