Workflow construction method and system

Abstract

The invention discloses a workflow construction method and system, and belongs to the technical field of electric digital data processing, and the method comprises the following steps: obtaining a language or text for constructing a workflow; obtaining an intention of the language or the text based on a large-scale language model; matching a component according to the intention, wherein the component is used for executing a task corresponding to the intention; judging whether interaction association exists between the two components or not; if yes, establishing a transmission rule for the two components; and if not, establishing a connection for the component, and obtaining a first workflow. On the basis of a large rule language model, the intention of a user is recognized, and on the basis, the adaptability, expansibility and automation of the workflow can be improved by matching corresponding components and organizing the components into the complete workflow; the method can quickly respond to the request of the user, and can be suitable for various different application scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a method and system for constructing a workflow. Background Art

[0002] The importance of process management and automated processing in modern enterprises is becoming increasingly prominent. Traditional process management systems usually rely on manual operations and preset rules, which are inefficient and easily affected by human factors. With the development of artificial intelligence (AI) technology, intelligent process management solutions have gradually become the focus of the market.

[0003] There are two workflow management technologies currently used: Business process management (BPM) systems mainly rely on process models and rule engines to perform automated tasks through predefined rules. Although such systems can improve efficiency to a certain extent, they lack adaptability and flexibility when faced with complex and dynamic business environments. In the face of ever-changing business needs and environments, the rule engines of traditional BPM systems need to be updated frequently, with high maintenance costs. They rely on preset rules, cannot autonomously learn and optimize processes, and have difficulty handling complex business logic.

[0004] Robotic process automation (RPA) technology simulates manual operations to perform repetitive tasks, reduce human errors and improve efficiency. However, the limitations of RPA technology are the lack of intelligent decision-making capabilities and poor adaptability to environmental changes. It can only perform predefined tasks and cannot make intelligent decisions based on real-time data; when faced with changing business processes, a large amount of customized development is required, which increases the complexity of implementation and maintenance.

[0005] Therefore, it is necessary to design a workflow construction method based on artificial intelligence and large-scale language model (LLM). Summary of the invention

[0006] In view of the above-mentioned technical problems existing in the prior art, the present invention provides a method and system for constructing a workflow, which understands the designer's intention based on a large-scale language model, and establishes a workflow according to the intention, thereby improving the adaptability, scalability and automation of the workflow.

[0007] The present invention discloses a method for constructing a workflow, comprising the following steps: obtaining a language or text for constructing a workflow; obtaining the intent of the language or text based on a large-scale language model; matching components according to the intent, the components being used to execute tasks corresponding to the intent; determining whether two components have an interactive association; if so, establishing a transfer rule for the two components; if not, establishing a connection for the components to obtain a first workflow.

[0008] Preferably, the method of matching components comprises: According to the text, obtain relevant knowledge content in the knowledge base; Based on a large-scale language model, the intent of the text and knowledge content is obtained, and entities are extracted; According to the intent and entity, and the intent category standard library, matching the corresponding intent type; According to the intent type, matching corresponding components; Obtaining the urgency of the intention according to the text, the urgency keyword thesaurus and the urgency classification threshold; Based on the urgency, the intent types / components are prioritized.

[0009] Preferably, the two components include a first component and a second component, and the method for the two components to interact includes: Determine whether the first task of the first component satisfies any of the following trigger conditions or a combination thereof: a trigger condition based on a business rule, a trigger condition based on a data state, and a trigger condition based on time; If satisfied, a subtask is established according to the first task; and the subtask is executed through the second component according to the intent type, urgency, processing status, and data transfer rules and data conversion mapping relationship.

[0010] Preferably, the method for establishing a connection for a component comprises: Obtaining a connection template, wherein the connection template includes: a business process template, an approval process template, and an emergency processing process template; Based on the intent and its complexity, select the corresponding connection template; Determine the order of components based on data dependencies; Connect each component according to the connection template and the order of components, and establish data transfer mapping between components; The data transfer mapping includes: unified data format, data type conversion between groups, data integrity verification, process status tracking, component execution status maintenance and abnormal status processing.

[0011] Preferably, the method for workflow optimization includes: inspecting, analyzing or periodically evaluating the first workflow; or Obtaining the question text of the first workflow; According to the intention of the question text and a preset question template, or the result of inspection, analysis or regular evaluation, the first workflow is adjusted to obtain a second workflow.

[0012] Preferably, the inspection, analysis and periodic evaluation include: Process integrity check, performance bottleneck identification, exception handling coverage, business rule adaptability assessment, processing efficiency data analysis, error rate statistical analysis, user satisfaction analysis, and resource utilization analysis; Based on the results of the inspection, analysis or periodic evaluation, components, parameters and first work flows are optimized.

[0013] Preferably, the problem text includes: process diagnosis problems, rule improvement problems and performance optimization problems. Methods for adjusting the first workflow according to the question text include: Collect question text based on any of the following methods or their combination: structured question templates, custom questions, option-based question confirmation; Based on the question text, obtain question intent and question parameters; Adjust workflow rules, processing parameters, and component configurations based on the problem intent and problem parameters.

[0014] Preferably, the method for executing the workflow includes: Get the user's input text; Based on a large-scale language model, the intent of the input text is obtained and the intent type is matched; Obtaining entity parameters of the input text and generating a first task; Obtaining the urgency of the input text, and setting the priority of the first task according to the urgency; According to the intent type, matching a corresponding first component; According to the priority, executing the first task by the first component; Determining whether the first task is interactively associated with the second component; If yes, generate a subtask according to the first task, and call the second component to execute the subtask; Generate reply language according to the first component and the situation of calling the second component; The reply language is returned to the user.

[0015] The present invention also provides a system for implementing the above method, comprising a first input module, an intention recognition module, a component matching module, an interaction module, and a connection module; The first input module is used to obtain the language or text for constructing the workflow; The intention recognition module is used to obtain the intention of the language or text based on a large-scale language model; The component matching module is used to match components according to the intent, and the components are used to execute the intent; The interaction module is used to establish a transfer rule for two components when there is an interactive association between the two components; The connection module is used to establish a connection for the components to obtain a first workflow.

[0016] Preferably, the system further comprises an optimization module, a second input module and a question-answering module; The optimization module is used to check, analyze or regularly evaluate the first workflow; or obtain the problem text of the first workflow; and adjusting the first workflow according to the intention of the question text and a preset question template, or the result of inspection, analysis or regular evaluation, to obtain a second workflow; The second input module is used to obtain the user's input text; The question-answering module is used to generate a reply language according to the input text and the workflow, and return the reply language to the user; The architecture of the system includes: workflow definition layer, task scheduling layer, parsing layer, engine layer and storage layer; The workflow definition layer includes a code module, a branch module, a loop module, an interaction module, an integration module, a knowledge base module, and an intent recognition module; The task scheduling layer is used to manage task planning and resource allocation of workflows; The parsing layer is used to analyze the workflow and build the logical relationship between modules; The engine layer includes a variable replacement module, a workflow execution engine, and an action processor; The storage layer includes a data model, a consistency check module, a backup module and a recovery module.

[0017] Compared with the prior art, the beneficial effects of the present invention are: based on the large rule language model, the user's intention is identified, and on this basis, by matching the corresponding components and organizing the components into a complete workflow, the adaptability, scalability and automation of the workflow can be improved; it can quickly respond to user requests and can be applicable to a variety of different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of a method for constructing a workflow of the present invention; Figure 2 It is a system logic block diagram of the workflow construction of the present invention; Figure 3 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The present invention is further described in detail below in conjunction with the accompanying drawings: The first aspect of the present invention provides a method for constructing a workflow. Figure 1 , including the following steps: Step 101: Obtain a language or text for constructing a workflow.

[0021] The user's input can be text or voice.

[0022] Step 102: Based on a large-scale language model (LLM), the intent of the language or text is obtained.

[0023] Step 103: Match one or more components according to the intent.

[0024] Components are used to perform tasks corresponding to the intent; components can also be described as nodes.

[0025] Step 104: Determine whether there is an interactive relationship between two components.

[0026] If yes, execute step 105 : establish transfer rules and mapping for the two components, and execute step 106 .

[0027] If not, execute step 106: establish connection and data transfer mapping for the components to obtain a first workflow.

[0028] Step 107: Check, analyze or regularly evaluate the first workflow.

[0029] Step 108: Obtain the question text of the first workflow.

[0030] Step 109: According to the intention of the question text and the preset question template, as well as the results of inspection, analysis or regular evaluation, the first workflow is adjusted to obtain a second workflow.

[0031] Based on the large rule language model, the user's intention is identified. On this basis, by matching the corresponding components and organizing the components into a complete workflow, the adaptability, scalability and automation of the workflow can be improved; it can quickly respond to user requests and can be applied to a variety of different application scenarios.

[0032] In step 103, the method of matching components includes: Step 301: Obtain relevant knowledge content in a knowledge base according to the text.

[0033] Knowledge content can be used to supplement the content of the text.

[0034] Step 302: Based on a large-scale language model, the intent of the text and knowledge content is obtained, and entities are extracted.

[0035] Step 303: Match the corresponding intent type according to the intent and entity, and the intent category standard library.

[0036] Step 304: Match corresponding components according to the intent type.

[0037] Step 305: Obtain the urgency of the intention according to the text, the emergency keyword vocabulary and the urgency classification threshold.

[0038] Step 306: Set a priority for the intent type / component based on the urgency.

[0039] In step 105, the two components include a first component and a second component, and the method for the two components to interact includes: Step 501: Determine whether the first task of the first component satisfies any one of the following trigger conditions or a combination thereof: trigger conditions based on business rules, data status, and time.

[0040] If so, execute step 502: create a subtask according to the first task; and call the second component to execute the subtask according to the intent type, urgency, processing status, data transfer rules and data conversion mapping relationship.

[0041] If not satisfied, only the first component is called to complete the first task.

[0042] In step 109, the inspection, analysis and periodic evaluation include: process integrity inspection, performance bottleneck identification, exception handling coverage, business rule adaptation evaluation, processing efficiency data analysis, error rate statistical analysis, user satisfaction analysis, and resource utilization analysis. Based on the results of the inspection, analysis or periodic evaluation, the components, parameters and the first workflow are optimized to achieve continuous updating and optimization of the workflow.

[0043] The problem text includes: process diagnosis problem, rule improvement problem and performance optimization problem. The method of adjusting the first workflow according to the problem text includes: Step 901: Collect question text based on any one of the following methods or a combination thereof: structured question template, custom question, option question confirmation.

[0044] Step 902: Based on the question text, obtain question intent and question parameters.

[0045] Step 903: According to the question intent and question parameters, adjust the workflow rules, processing parameters and component configuration.

[0046] The established first workflow may be optimized by natural language input, but is not limited thereto, and the first workflow may also be optimized manually.

[0047] The methods of workflow execution include: Step 921: Obtain the user's input text.

[0048] Step 922: Based on the large-scale language model, obtain the intent of the input text and match the intent type.

[0049] Step 923: Obtain entity parameters of the input text and generate a first task.

[0050] Step 924: Obtain the urgency of the input text, and set the priority of the first task according to the urgency.

[0051] Step 925: Match the corresponding first component according to the intent type.

[0052] Step 926: Execute the first task through the first component according to the priority.

[0053] Step 927: Determine whether the first task is interactively associated with the second component.

[0054] If so, execute step 928: generate a subtask according to the first task, and call the second component to execute the subtask.

[0055] If not, execute step 929.

[0056] Step 929: Generate a reply language according to the first component (and the situation of calling the second component). Preferably, the reply language is a natural language.

[0057] Step 930: Return the reply language to the user.

[0058] A second aspect of the present invention provides a system for implementing the above method, such as Figure 2 As shown, it includes a first input module 1, an intention recognition module 2, a component matching module 3, an interaction module 4, and a connection module 5; The first input module 1 is used to obtain the language or text for constructing the workflow; The intention recognition module 2 is used to obtain the intention of the language or text based on a large-scale language model; The component matching module 3 is used to match components according to the intention, and the components are used to execute the intention; The interaction module 4 is used to establish a transfer rule for the two components when there is an interactive association between the two components; The connection module 5 is used to establish a connection for the components to obtain a first workflow.

[0059] More specifically, the system further comprises an optimization module 6, a second input module 7 and a question-answering module 8; The optimization module 6 is used to check, analyze or regularly evaluate the first workflow; or obtain the problem text of the first workflow; and adjusting the first workflow according to the intention of the question text and a preset question template, or the result of inspection, analysis or regular evaluation, to obtain a second workflow; The second input module 7 is used to obtain the user's input text; The question-answering module 8 is used to generate a reply language according to the input text and the workflow, and return the reply language to the user.

[0060] like Figure 3 The architecture of the system includes: a workflow definition layer 11, a task scheduling layer 12, a parsing layer 13, an engine layer 14 and a storage layer 15.

[0061] The workflow definition layer 11 includes a code module, a branch module, a loop module, an interaction module, an integration module, a knowledge base module, an intent recognition module and an intent classification module.

[0062] Code module (Code): used to define the code snippets that need to be executed, currently supporting the commonly used Python / Javascript languages. Branch module Branch: implements conditional branch logic. Loop module Loop: processes loop structures. Interaction module Webhook: HTTP-based callback function, allowing lightweight, event-driven communication between two application programming interfaces (APIs) for interaction with external systems. Integration module Integration: integrates external services and APIs. Knowledge base module Knowledge: based on large model vectorization to store user knowledge data, retrieve text content related to user questions from the knowledge base, and can be used as context for downstream LLM and intent classification. Intent recognition module Llm: calls advanced large language models to answer questions or process user natural language. Intent classification module Intent: LLM intent recognition classification, by defining classification descriptions, the question classifier can infer the matching classification based on user input and output the classification results.

[0063] In a specific embodiment, the workflow structure is defined using JSON format. The definition code is as follows: { "id": "d60ea71b-84a0-98ef-5546-f7ba724d9d8d", "name": "startTrigger", "actionType": "TRIGGER", "type": "WEBHOOK", "layout": { "position": { "x": "223", "y": "323" }, "width": "111", "hight": "222" }, "data": { "type": "", "inputs": { "p1": "{{question}}", "p2": "{{step_13}}" }}} Among them, id: This is the unique identifier of the node / component. name: The name of the node, indicating that this is a start trigger. actionType: The action type of the node. type: "WEBHOOK" - The node type is Webhook, which means that it triggers the workflow through HTTP requests. layout: The layout information of the node on the interface. position indicates the horizontal and vertical position of the node on the interface; width and height represent the width and height of the node respectively; data: The user-configured data information stored by the node. type: The type of the node. inputs: Defines the input parameters of the node.

[0064] The workflow definition layer can also provide a graphical interface, allowing users to design workflows by dragging and dropping, and finally arranging the data of each node into a complete workflow data protocol. The workflow definition layer can implement a version control system and support workflow history tracking and rollback.

[0065] The task scheduling layer 12 is used to manage task planning and resource allocation of workflows.

[0066] The main functions include: once job: a task that is executed once. repeating job: a task that is executed repeatedly periodically. Implementation method: Use the distributed task scheduling framework PowerJob to manage task execution. Implement priority queues to ensure that key tasks are executed first. Use dynamic resource allocation algorithms to automatically adjust task execution resources according to system load.

[0067] The parsing layer 13 is used to analyze the workflow and build the logical relationship between components. The core module is ParseNodeRelationService, which is responsible for analyzing the workflow definition and building the logical relationship between nodes. Implementation method: Use directed acyclic graph (DAG) to represent node relationships. Implement topological sorting algorithm to determine the node execution order. Build a dependency matrix to optimize parallel execution strategy.

[0068] The engine layer 14 includes a variable replacement module, a workflow execution engine, and an action processor. Functions of the variable replacement module: dynamically replace variables filled in by users in the workflow, replace variables with real data before node execution, and ensure that data is correctly transmitted between nodes. Implementation: Customize the parsing service VariableService, use regular expressions for pattern matching and replacement, and support complex nested variable structures.

[0069] The workflow execution engine FlowExecutor is the main execution engine of the workflow. Components: a) ExecutionState: maintains the workflow execution state and stores key information during the workflow execution process. b) ExecutionOutput: manages the output of execution results.

[0070] Action processors include a variety of specialized processors, such as: IntegrationActionHandler: handles integration actions. BranchActionHandler: handles branch logic. CodeActionHandler: executes code snippets. LoopActionHandler: manages loop operations. KnowledgeActionHandler: handles LLM vectorized knowledge base operations. LlmActionHandler: handles LLM natural language dialogue operations. IntentActionHandler: manages intelligent recognition of user intent and intelligent distribution of data flows.

[0071] The engine layer 14 also includes multiple functional modules: CODE: executes custom code. Branch: implements conditional branching. Loop: implements loop logic. Knowledge: vector retrieval, users quickly find user structured data. LLM: integrates a large language model for natural language processing tasks. Intent: inherits the large language model, realizes the model automatically recognizes user intent, and automatically distributes tasks.

[0072] Implementation method: Use the responsibility chain model to design action processors to improve the scalability of the system. Use the state machine model to manage the workflow execution state to ensure the consistency of state transitions. Implement event-driven architecture to improve system responsiveness and concurrent processing capabilities. Integrate monitoring and logging systems to track workflow execution in real time.

[0073] The storage layer 15 includes a data model, a consistency check module, a backup module, and a recovery module. PostgreSQL is used as the main data storage system. Implementation method: Design an optimized data model to support efficient workflow data storage and retrieval. Implement a partition table strategy to improve large-scale data processing performance. Use a transaction mechanism to ensure data consistency. Implement a regular backup and recovery mechanism to ensure data security.

[0074] The present invention supports intelligent task decomposition, dynamic workflow generation, intelligent error handling and self-healing, and performance optimization. Intelligent task decomposition: Use the LLM model to automatically decompose complex tasks into executable subtasks; implement an adaptive task granularity adjustment algorithm to dynamically optimize the task decomposition strategy based on the execution effect. Dynamic workflow generation: Based on historical execution data and current context, use machine learning algorithms to dynamically generate and optimize workflows; implement a workflow template library to support rapid reuse and customization. Intelligent error handling and self-healing: Design an error detection system based on a rule engine; implement an automatic retry mechanism, and intelligently select a recovery strategy based on the error type. Performance optimization: The workflow preheating mechanism can be used to load common components in advance; use caching technology to optimize data access performance; implement an intelligent resource scheduling algorithm to dynamically allocate computing resources based on node complexity.

[0075] The present invention uses natural language to generate workflows, allowing users to create and modify workflows using natural language descriptions, greatly reducing the user's usage threshold and improving the accessibility of the system. It includes natural language understanding, workflow mapping, context understanding and maintenance, natural language generation, workflow optimization suggestions, interactive workflow tuning, and interactive workflow tuning.

[0076] Natural language understanding: parse the user's natural language input and extract key information and intent. Implementation method: abstract workflow node data model; use advanced natural language processing model GPT to learn data model; user intent recognition, accurately capture user goals and needs; the model generates node data based on the information provided by the user. Key parameters: model type: gpt-4o (for fast processing); intent classification accuracy: >95%; entity recognition F1 score: >0.9.

[0077] Workflow mapping: Map the results of natural language understanding to specific workflow structures and components. Implementation method: Design a mapping rule base from semantics to workflow components; implement a reasoning mechanism based on node relationships to handle complex logical relationships; use models to evaluate the accuracy of node data and node relationship mapping. Key technologies: Component relationship construction: Use Postgres to store workflow components and relationships; Reasoning algorithm: Use rule-based reasoning methods; Mapping accuracy: >90%.

[0078] Context understanding and maintenance: Maintain user intent and workflow context in multiple rounds of dialogue. Implementation method: Implement a session state tracker to record the user's historical input and system response. Use prompts to solve the user context hallucination problem. Key features: Context window size: the last 5 rounds of dialogue, 128k information storage; Conflict resolution accuracy: >85%.

[0079] Natural language generation: Convert system operations and feedback into natural language to interact with users. Implementation method: Use template-based generation methods to ensure output consistency; combine with GPT models to provide more natural and fluent responses; implement multi-round dialogue management to maintain dialogue coherence and context understanding. Key parameters: Response generation time: <500ms; User satisfaction (assessed through feedback): >4.5 / 5.

[0080] Workflow optimization suggestions: Provide workflow optimization suggestions based on the user's natural language description. Implementation method: Analyze the task objectives and constraints described by the user; use historical workflow data and the GPT model to generate optimization suggestions. Key indicators: Suggestion adoption rate: >70%, workflow execution efficiency improvement: average >20%.

[0081] Interactive workflow tuning: Allow users to gradually adjust and optimize workflows through natural language dialogue. Implementation method: Design an interactive question-and-answer module to guide users to clarify various aspects of the workflow; implement real-time workflow visualization to intuitively display the effects of user adjustments; use incremental learning algorithms to continuously optimize the system's understanding of user preferences. Key features: Average dialogue rounds (to complete a workflow optimization): <5 rounds, user operations reduced (compared to traditional interfaces): >60%.

[0082] Interactive workflow tuning allows users to gradually adjust and optimize workflows through natural language dialogue. Implementation method: Design an interactive question-and-answer module to guide users to clarify various aspects of the workflow; implement real-time workflow visualization to intuitively display the effects of user adjustments; use incremental learning algorithms to continuously optimize the system's understanding of user preferences. Key features: Average dialogue rounds (to complete a workflow optimization): <5 rounds, user operations reduced (compared to traditional interfaces): >60%.

[0083] The present invention has a high degree of integration and synergy, and realizes seamless integration of natural language interface and workflow definition layer. The result processing of natural language understanding is added to the parsing layer (ParseNodeRelationService). In the extended FlowExecutor, dynamic workflow execution based on natural language generation is supported.

[0084] The present invention also implements an intention confirmation mechanism to prevent misoperation, sets permission control to limit the scope of operations performed by specific users through natural language, and implements sensitive information filtering to ensure that confidential data is not leaked during natural language interaction.

[0085] In a specific embodiment, an intelligent HR customer service workflow is constructed.

[0086] Objective: Create an intelligent HR customer service workflow to automatically handle employees' common inquiries, leave applications, and document requests, thereby improving the efficiency of the HR department and employee satisfaction.

[0087] Here are the steps: Step 951, user input: User (HRBP) input: "Create an HR customer service workflow that can answer common employee questions, process leave requests, provide company policy documents, and transfer to manual HR if the question is complex. The workflow should be able to identify emergency situations and prioritize them." Step 952, natural language understanding: use the GPT model to perform intent recognition and entity extraction; match related components based on the information. Key components identified: question-answering system, vacation processing, social security consultation, document provision, emergency identification, and manual transfer. Intent classification accuracy: 98.2%.

[0088] The component matching process after intent recognition is as follows: preset the second input module to process employee input requests; configure the request format recognition rules; set the request priority judgment parameters. Match LLM for text analysis, configure the emergency keyword vocabulary; set the urgency classification threshold.

[0089] Intent classification: Use the CODE component to implement classification logic, establish a standard library for intent categories, and configure the intent recognition confidence threshold. Knowledge base query node, connect to the Knowledge component, set knowledge retrieval rules, and configure similarity matching parameters.

[0090] Step 953: Automatically generate workflow.

[0091] Interaction between components: Data flow design, define standard data structure (including intent type, urgency, processing status, etc.); set data transmission rules; configure data conversion mapping relationship. Trigger condition setting: conditional triggering based on business rules, conditional triggering based on data status, and conditional triggering based on time.

[0092] Automated connection implementation: Preset connection templates, including common HR business process templates, standard approval process templates, and emergency processing process templates. Dynamic connection generation: Automatically select connection templates based on intent complexity, determine node order based on data dependencies, and support automatic generation of conditional branches.

[0093] Data transfer mechanism: data transfer mapping between nodes: transfer association between nodes / components; unified data format definition; automatic data type conversion; data integrity verification. State management: process state tracking; node execution state maintenance; abnormal state processing.

[0094] Step 954: Interactive optimization design.

[0095] System optimization suggestion generation design: analysis based on rule engine: process integrity check rules; performance bottleneck identification rules; exception handling coverage rules; business rule adaptability evaluation; analysis based on historical data; processing efficiency data analysis; error rate statistical analysis; user satisfaction analysis; resource utilization analysis.

[0096] User-guided optimization design: process diagnosis questions, such as "Is the average processing time of this node reasonable?", "Is it necessary to add additional review links?", "Does the existing emergency processing mechanism meet the needs?". Rule improvement questions: For example, "Please confirm the time threshold for vacation approval"; "Is it necessary to set up differentiated processes for different departments?", "How to define the priority of emergency situations?". Performance optimization questions, such as "Is it necessary to process multiple approval links in parallel?", "What automated processing rules need to be set?", "How to improve the efficiency of knowledge base query?".

[0097] Optimize feedback collection and application: Question collection method: structured question template; custom question input; optional question confirmation. Feedback application mechanism: automatic update of process rules; dynamic adjustment of processing parameters; optimization of component configuration.

[0098] Continuous optimization mechanism: Regular optimization evaluation: performance indicator analysis; user satisfaction evaluation; process efficiency evaluation. Dynamic adjustment strategy: automatic optimization based on evaluation results; rule update based on feedback; support manual intervention adjustment.

[0099] The established workflow performance parameter settings are as follows: 5. LLM uses the GPT-4O model with a response time limit of 3 seconds; the knowledge base uses Milvus, and the index update frequency is once an hour; the intent classification uses a fine-tuned GPT model, and the classification threshold is set to 0.85; the emergency identification priority setting is: security issues > discrimination / harassment complaints > salary issues; the maximum concurrent processing number of the workflow: 200 requests / minute.

[0100] Test execution: 5,000 real employee request data were used for testing (including inquiries, leave applications, document requests, etc.), average first response time: 2.3 seconds, problem resolution rate (without human intervention): 87%, automatic processing rate of leave applications: 95%, document provision accuracy: 99%, emergency identification accuracy: 99.5%, employee satisfaction score (based on feedback): 4.6 / 5.

[0101] The execution process of a specific request is as follows: a. The employee inputs: "I want to apply for a vacation from next Wednesday to Friday."; b. The emergency identification node determines it as non-emergency (confidence 99.8%); c. The intent classification node identifies it as "vacation application" (confidence 97.5%); d. The vacation application processing node executes: check the employee's vacation balance (15 days of annual leave remaining); confirm the number of days applied for (3 days); trigger the supervisor approval process; e. The response generation node creates a reply: "Your 3-day vacation application has been received and submitted to your supervisor for approval. Your current annual leave balance is 15 days. We will notify you of the approval result by email within 24 hours. If you have any questions, please feel free to ask."; f. The reply is sent to the employee via Webhook.

[0102] System learning and optimization: automatically recommend relevant information to HR administrators and add it to the knowledge base; optimize the leave application process and add automatic calendar synchronization function.

[0103] The intelligent HR customer service workflow system demonstrated excellent performance and efficiency. It successfully automated 87% of employee requests, significantly reducing the workload of the HR team. The system was able to quickly respond to employee inquiries, accurately process leave applications, and effectively identify and prioritize emergency situations. 95% of leave applications were automated, greatly improving efficiency.

[0104] It is particularly noteworthy that the learning ability of the system of the present invention enables it to continuously optimize and expand its functions. For example, it automatically identifies new topics that employees are concerned about (such as high temperature subsidy policies) and actively recommends updating the knowledge base, which greatly improves the practicality and adaptability of the system.

[0105] The employee satisfaction score reached 4.6 / 5, reflecting that the system not only improved the work efficiency of the HR department, but also significantly improved the employee experience. The system's fast response time (average 2.3 seconds) and high accuracy were generally recognized by employees.

[0106] The present invention has the following features: Intelligent process modeling: Automatic conversion of natural language description to process model is realized through NLP technology. AI-driven task scheduling: Dynamic optimization of task allocation and resource scheduling using machine learning algorithms. Adaptive rule engine: Integrated LLM, the system autonomously learns and optimizes business rules to improve decision-making accuracy. Real-time monitoring and data analysis: Provides real-time visual monitoring and data analysis functions to optimize process execution. Multi-platform integration: Standardized API interface realizes seamless connection with various enterprise systems.

[0107] Enterprises will see significant improvements in business process management: Efficiency improvement: Intelligent task scheduling and resource allocation optimization to achieve efficient operation of business processes. Intelligent decision-making: The adaptive rule engine automatically optimizes business rules through AI technology to improve the accuracy and timeliness of decisions. Cost reduction: Reduce manual intervention and maintenance costs, and achieve process automation through intelligent means. Enhanced flexibility: The system can adapt to business changes, respond quickly to dynamic needs, and improve the company's resilience. Data-driven optimization: Through real-time monitoring and data analysis, continuously optimize business processes and improve overall operational efficiency.

[0108] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a workflow, characterized in that: The following steps are involved: Step 101: Obtaining a language or text for constructing a workflow; Step 102: Obtaining the intent of the language or text based on a large-scale language model; Step 103: matching a component according to the intent, the component being used to execute a task corresponding to the intent; Step 104: Determine whether there is an interactive association between two components; If yes, execute step 105: establish a transfer rule for the two components, and execute step 106; If not, execute step 106: establish a connection for the component to obtain a first workflow.

2. The method according to claim 1, characterized in that The methods for matching components include: According to the text, relevant knowledge content is obtained in a knowledge base; Based on a large-scale language model, the intent of the text and knowledge content is obtained, and entities are extracted; According to the intent and entity, and the intent category standard library, matching the corresponding intent type; According to the intent type, matching corresponding components; Obtaining the urgency of the intention according to the text, the urgency keyword thesaurus and the urgency classification threshold; Based on the urgency, the intent types / components are prioritized.

3. The method according to claim 1, characterized in that The two components include a first component and a second component, and the method for the two components to interact includes: Determine whether the first task of the first component satisfies any of the following trigger conditions or a combination thereof: a trigger condition based on a business rule, a trigger condition based on a data state, and a trigger condition based on time; If satisfied, a subtask is established according to the first task; and the subtask is executed through the second component according to the intent type, urgency, processing status, and data transfer rules and data conversion mapping relationship.

4. The method according to claim 1, characterized in that: Methods for establishing connections for components include: Obtaining a connection template, wherein the connection template includes: a business process template, an approval process template, and an emergency processing process template; Based on the intent and its complexity, select the corresponding connection template; Determine the order of components based on data dependencies; Connect each component according to the order of connection template and components, and establish data transfer mapping between components; The data transfer mapping also includes: unified data format, data type conversion between groups, data integrity verification, process status tracking, component execution status maintenance and abnormal status processing.

5. The method according to claim 1, characterized in that Also includes methods for workflow optimization: inspecting, analyzing or periodically evaluating the first workflow; or Obtaining the question text of the first workflow; According to the intention of the question text and a preset question template, or the result of inspection, analysis or regular evaluation, the first workflow is adjusted to obtain a second workflow.

6. The method according to claim 5, characterized in that Such inspections, analyses and periodic assessments include: Process integrity check, performance bottleneck identification, exception handling coverage, business rule adaptability assessment, processing efficiency data analysis, error rate statistical analysis, user satisfaction analysis, and resource utilization analysis; Based on the results of the inspection, analysis or periodic evaluation, components, parameters and first work flows are optimized.

7. The method according to claim 5, characterized in that The problem texts include: process diagnosis problems, rule improvement problems and performance optimization problems. Methods for adjusting the first workflow according to the question text include: Collect question text based on any of the following methods or their combination: structured question templates, custom questions, option-based question confirmation; Based on the question text, obtain question intent and question parameters; Adjust workflow rules, processing parameters, and component configurations based on the problem intent and problem parameters.

8. The method according to claim 1, characterized in that The methods of workflow execution include: Get the user's input text; Based on a large-scale language model, the intent of the input text is obtained and the intent type is matched; Obtaining entity parameters of the input text and generating a first task; Obtaining the urgency of the input text, and setting the priority of the first task according to the urgency; According to the intent type, matching a corresponding first component; According to the priority, executing the first task by the first component; Determining whether the first task is interactively associated with the second component; If yes, generate a subtask according to the first task, and call the second component to execute the subtask; Generate reply language according to the first component and the situation of calling the second component; The reply language is returned to the user.

9. A system, characterized in that: Used to implement the method according to any one of claims 1 to 8, the system comprises a first input module, an intention recognition module, a component matching module, an interaction module, and a connection module; The first input module is used to obtain the language or text for constructing the workflow; The intention recognition module is used to obtain the intention of the language or text based on a large-scale language model; The component matching module is used to match components according to the intent, and the components are used to execute the intent; The interaction module is used to establish a transfer rule for two components when there is an interactive association between the two components; The connection module is used to establish a connection for the components to obtain a first workflow.

10. The system according to claim 9, characterized in that Also includes an optimization module, a second input module and a question-answering module; The optimization module is used to check, analyze or regularly evaluate the first workflow; or obtain the problem text of the first workflow; and adjusting the first workflow according to the intention of the question text and a preset question template, or the result of inspection, analysis or regular evaluation, to obtain a second workflow; The second input module is used to obtain the user's input text; The question-answering module is used to generate a reply language according to the input text and the workflow, and return the reply language to the user; The architecture of the system includes: workflow definition layer, task scheduling layer, parsing layer, engine layer and storage layer; The workflow definition layer includes a code module, a branch module, a loop module, an interaction module, an integration module, a knowledge base module, and an intent recognition module; The task scheduling layer is used to manage task planning and resource allocation of workflows; The parsing layer is used to analyze the workflow and build the logical relationship between modules; The engine layer includes a variable replacement module, a workflow execution engine, and an action processor; The storage layer includes a data model, a consistency check module, a backup module and a recovery module.

Images