A workflow-based processing system

Abstract

This application discloses a workflow-based processing system. A scheduling execution server acquires first language instructions, decomposes these instructions to obtain first scenario requirements, and sends these requirements to a planning server. The planning server processes these requirements to obtain first node information, which is then sent to an orchestration server. The first node information includes at least the type and quantity of first nodes. The orchestration server fills in edge parameters based on the first node information to obtain a first executable result, which is then sent to a validator. The validator performs verification based on the first executable result to obtain a verification result. In other words, by processing the first language instructions through a scheduling execution server, planning server, orchestration server, and validator, a workflow is accurately and automatically generated, avoiding the consumption of manual resources and costs.

Description

Technical Field

[0001] This application relates to artificial intelligence, and to, but is not limited to, a workflow-based processing system. Background Technology

[0002] Workflows are systems that orchestrate LLMs and tools through predefined code paths, integrating various components in a specific logical order. Like an assembly line, they break down complex tasks into multiple simple steps, each with a clear goal and process, thereby completing a specific task. However, in real-world business scenarios, building traditional workflows often requires business knowledge and specialized skills.

[0003] With existing technology, the reliability of the generated workflow results relies on manual verification, which cannot fully guarantee their accuracy and increases the workload and cost of manual review. Summary of the Invention

[0004] In view of this, embodiments of this application provide a workflow-based processing system.

[0005] The technical solution of this application embodiment is implemented as follows: This application provides a workflow-based processing system, comprising: a scheduling execution server, a planning server, an orchestration server, and a validator; the scheduling execution server is configured to acquire a first language instruction, decompose the first language instruction to obtain a first scenario requirement, and send the first scenario requirement to the planning server; the planning server is configured to process based on the first scenario requirement to obtain first node information, and send the first node information to the orchestration server, wherein the first node information includes at least: a first node type and a first node quantity; the orchestration server is configured to fill edge parameters based on the first node information to obtain a first executable result, and send the first executable result to the validator; the validator is configured to perform verification based on the first executable result to obtain a verification result.

[0006] Optionally, the planning server includes: a node filtering unit; the node filtering unit is used to filter tasks in the node description template according to the requirements of the first scenario to obtain the first node type.

[0007] Optionally, the planning server further includes: a component filtering unit; the component filtering unit is used to filter tasks according to the requirements of the first scenario to obtain the number of the first nodes.

[0008] Optionally, the component filtering unit includes: a first calling module; the first calling module is used to obtain a first calling tool according to the first scenario requirements, and to perform task filtering on the first scenario requirements through the first calling tool to obtain the first number of nodes.

[0009] Optionally, the planning server further includes: a first reflection unit; the first reflection unit is used to reflect based on the first node information, obtain a first reflection result, and determine whether to make adjustments based on the first reflection result.

[0010] Optionally, the orchestration server includes: a structure orchestration unit and an edge module; the structure orchestration unit is used to fill the first node information into the edge module to obtain a first parameter template.

[0011] Optionally, the structure orchestration unit includes a second invocation unit; the second invocation unit is used to obtain a second invocation tool based on the first node information, and to fill the first node information into the edge module through the second invocation tool.

[0012] Optionally, the structure arrangement unit further includes: a connection module; the connection module is used to obtain a first number of connections based on the first node information, thereby performing node connection.

[0013] Optionally, the orchestration server further includes: a parameter filling unit and a node module; the parameter filling unit is used to concatenate the first parameter template with the node module to obtain a first executable result.

[0014] Optionally, the orchestration server further includes: a second reflection unit; the second reflection unit is used to reflect based on the first parameter template, obtain a second reflection result, and determine whether to make adjustments based on the second reflection result.

[0015] This application provides a workflow-based processing system. A scheduling execution server acquires first language instructions, decomposes them to obtain first scenario requirements, and sends these requirements to a planning server. The planning server processes these requirements to obtain first node information, which is then sent to an orchestration server. The first node information includes at least one type and one quantity of nodes. The orchestration server fills in edge parameters based on the first node information to obtain a first executable result, which is then sent to a validator. The validator performs verification based on the first executable result to obtain a verification result. In other words, by processing the first language instructions through a scheduling execution server, planning server, orchestration server, and validator, a precise and automated workflow generation is achieved, avoiding the consumption of manual resources and costs. Attached Figure Description

[0016] Figure 1 A schematic diagram of a workflow-based processing system provided in an embodiment of this application; Figure 2 Another flowchart of a workflow-based processing system provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] Please refer to Figure 1 , Figure 2 ,in, Figure 1 This application provides a schematic diagram of an implementation of a workflow-based processing system, which includes: a scheduling execution server, a planning server, an orchestration server, and a validator. The scheduling execution server is used to acquire first language instructions, decompose the first language instructions to obtain first scenario requirements, and send the first scenario requirements to the planning server; the planning server is used to process based on the first scenario requirements, obtain first node information, and send the first node information to the orchestration server, wherein the first node information includes at least: first node type and first node quantity; the orchestration server is used to fill edge parameters based on the first node information, obtain a first executable result, and send the first executable result to the verifier; the verifier is used to perform verification based on the first executable result and obtain a verification result.

[0019] The scheduling execution server can be a Scheduler Agent, responsible for three tasks: task decomposition, execution scheduling, and result reflection; the planning server can be a Filter Agent, responsible for filtering task planning, tools invocation, and result reflection; the orchestration server can be an Orchestrator Agent, responsible for orchestrating task planning, tools invocation, and result reflection; and the validator can be an Execution Engine, responsible for verifying the correctness of edge parameters and node parameters during the generation process.

[0020] Specifically, 5.1 The Scheduler Agent The SchedulerAgent is responsible for three tasks: task breakdown, execution scheduling, and result evaluation. First, the SchedulerAgent dynamically breaks down the task based on user input. For example, if the user provides a specific business scenario requirement, the Agent first calls the Filter Agent to confirm the required node types and quantities, then calls the Orchestrator Agent to fill in the parameters for subsequent nodes and edges. Simultaneously, the SchedulerAgent checks the results returned by the FilterAgent and Orchestrator Agent. If the requirements for the current scenario cannot be met, the SchedulerAgent sends the feedback to the corresponding Agent and regenerates the task. Only when the SchedulerAgent confirms that the target task has been completed will it call the next Agent for execution. If the user's requirement is simply to modify parameters or confirm node types, the SchedulerAgent will only call the Orchestrator Agent or Filter Agent to execute the task individually. This flexible scheduling method avoids wasting resources and time, resulting in more efficient task completion.

[0021] 5.2 The Filter Agent The Filter Agent is responsible for three parts: task planning, tool invocation, and result reflection. Task planning combines user input and the Scheduler Agent's analysis to determine which tool to use to complete the task. Similarly, the Filter Agent dynamically plans based on input information and reflects on the tool's returned results to determine whether the corresponding tool needs to be invoked again.

[0022] 5.2.1 Node Filtering Tool The node filtering tool selects task-related nodes from all node description templates based on user input and feedback from the Filter Agent. These templates contain descriptions for over 15 node types, including code execution, classifiers, large models, if-else statements, loop iterations, knowledge bases, MCP tools, and parameter extraction. This invention uses DeepSeek-R1-0528 for input understanding and node type matching to ensure the reliability of subsequent workflow construction.

[0023] 5.2.2 Component Quantity Filtering Tool The component quantity filtering tool requires user input, feedback from the Filter Agent, and data on the types of nodes to be filtered by the node filtering tool as input. In complex business scenarios, multiple LLM nodes, MCP tool nodes, or nodes with multiple branches such as if-else statements and classifiers may need to be used simultaneously, significantly increasing the complexity of workflow setup. Therefore, the component quantity filtering tool determines the number of nodes provided by the node filtering tool based on task requirements, configuring a unique ID for each node. These node IDs will be used in subsequent edge connections.

[0024] 5.3 The Orchestrator Agent The Orchestrator Agent is responsible for three parts: task planning, tool invocation, and result reflection. Task planning combines user input, the Scheduler Agent's analysis, and the Filter Agent's output to determine which tool to use to complete the task. The Orchestrator Agent also dynamically plans based on input information and reflects on the tool's returned results to determine if the corresponding tool needs to be invoked again.

[0025] 5.3.1 Structure Arrangement Tools The orchestration tool is crucial in the workflow setup process, responsible for branching, connecting nodes, and filling edge parameters. It first understands the input instructions comprised of user requirements and the OrchestratorAgent's considerations. Then, it fills in the previously generated node types and quantities into a specified blank workflow parameter template, requiring the orchestration tool to populate the edge parameters. The workflow parameter template is divided into Nodes and Edges modules. The orchestration tool only needs to fill in the parameters in the Edges module. Each edge in Edges requires filling in a Source (source node) and a Target (target node). This also necessitates the tool determining the total number of edges needed for the entire task to connect each node.

[0026] 5.3.2 Parameter Filling Tool The parameter filling tool is the final step in orchestrating the entire workflow, outputting the final executable workflow graph structure. At this stage, user requirements, the Orchestrator Agent's thought process, the output of the structure orchestration tool, blank parameter templates for the required nodes, and parameter description templates are all required by the tool. The parameter filling tool then appends the filled node parameter templates to the Nodes module in the workflow parameter template. Due to the large number and variety of node types involved in the task, the number of node parameters to be filled is substantial. Therefore, for large model nodes, intent recognition nodes, and classifier nodes, the model temperature (0.3), model type (using the DeepSeek-R1 model uniformly), memory capacity (3 rounds), and maximum output tokens (4096) are fixed. For MCP services, the API keys for each MCP service in the MCP node are fixed in advance. For knowledge base nodes, the recall strategy (mixed weight mode, recall quantity 3, threshold 0.5) is fixed.

[0027] 5.3 Execution Engine The Execution Engine is the execution engine that validates the generated workflow graph, ensuring the correctness of edge and node parameters during the generation process. If execution fails, the Execution Engine identifies the cause of the failure and outputs it to the Scheduler Agent for subsequent modifications.

[0028] The method in this embodiment may include a planning server comprising: a node filtering unit; the node filtering unit is used to filter tasks in a node description template according to the requirements of a first scenario to obtain a first node type.

[0029] The method in this embodiment may include: the planning server further includes: a component filtering unit; the component filtering unit is used to filter tasks according to the requirements of the first scenario to obtain the number of first nodes.

[0030] The method in this embodiment may include: the component filtering unit includes: a first calling module; the first calling module is used to obtain a first calling tool according to the first scenario requirements, and to perform task filtering on the first scenario requirements through the first calling tool to obtain the first number of nodes.

[0031] Specifically, the task planning process combines user input requirements with analysis of the scheduling execution server to determine which tool to call to complete the task, meaning the task is completed through the first tool called.

[0032] The method in this embodiment may include: the planning server further includes: a first reflection unit; the first reflection unit is used to reflect based on the first node information, obtain a first reflection result, and determine whether to make an adjustment based on the first reflection result.

[0033] Specifically, the first reflection unit will dynamically plan based on the input information, and at the same time, reflect on the results returned by the reflection tool to determine whether it is necessary to call the corresponding tool again.

[0034] The method in this embodiment may include: an orchestration server comprising: a structure orchestration unit and an edge module; the structure orchestration unit being used to fill the edge module with first node information to obtain a first parameter template.

[0035] The method in this embodiment may include: the structure orchestration unit includes: a second calling unit; the second calling unit is used to obtain a second calling tool according to the first node information, and fill the first node information into the edge module through the second calling tool.

[0036] Specifically, the task planning process combines user input requirements, Scheduler Agent's analysis and considerations, and The Filter Agent's output results to determine which tool to call to complete the task, that is, to complete the task through a second calling tool.

[0037] The method in this embodiment may include: the structure arrangement unit further includes: a connection module; the connection module is used to obtain a first number of connections based on the first node information, thereby performing node connection.

[0038] The method in this embodiment may include: the orchestration server further includes: a parameter filling unit and a node module; the parameter filling unit is used to concatenate the first parameter template with the node module to obtain a first executable result.

[0039] The method in this embodiment may include: the orchestration server further includes: a second reflection unit; the second reflection unit is used to reflect based on the first parameter template, obtain a second reflection result, and determine whether to make an adjustment based on the second reflection result.

[0040] Specifically, the second reflection result will be dynamically planned based on the input information, and the results returned by the reflection tool will be used to determine whether the corresponding tool needs to be called again.

[0041] The method in this embodiment may include: a multi-agent automatic workflow optimization framework - MAFlow, which breaks down complex logical tasks into multiple agents working together to complete them. MAFlow consists of four parts: Scheduler Agent, FilterAgent, Orchestrator Agent, and Execution Engine, and the overall framework is as follows: Figure 1 As shown, when the SchedulerAgent receives a user request, it breaks down the task and calls the corresponding Filter Agent and Orchestrator Agent to complete it based on the execution logic. The Filter Agent and Orchestrator Agent then select appropriate tools to complete the task step by step. To verify the quality of the final generated workflow, the Execution Engine executes the generated workflow graph and performs dataset evaluation.

[0042] The advantages of this application compared to existing technologies are: (1) It has greatly improved the complexity of workflow construction and the richness of nodes. It can generate complex workflows with 10+ node types, 20+ parameters, and multiple branches, and can excellently meet most business scenario requirements.

[0043] (2) The designed multi-agent framework is no longer limited to specific tasks like a single agent. It can break down multiple complex tasks and call different agents to collaborate and complete them together. This helps to increase versatility, making it suitable for enterprises to use in a wider range of application scenarios, and enhancing the system's ability and adaptability to handle diverse and highly complex tasks.

[0044] (3) Adding result reflection and automated result verification to each agent helps to achieve dynamic optimization and higher-level intelligent decision-making in the workflow execution process, thereby effectively reducing the cost of manual intervention and ensuring the accuracy and availability of results.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A workflow-based processing system, characterized in that, The system includes: a scheduling execution server, a planning server, an orchestration server, and a verifier; The scheduling execution server is used to obtain a first language instruction, decompose the first language instruction to obtain a first scenario requirement, and send the first scenario requirement to the planning server; The planning server is used to process the requirements of the first scenario, obtain the first node information, and send the first node information to the orchestration server. The first node information includes at least: the type of first node and the number of first nodes. The orchestration server is used to fill in the edge parameters based on the first node information, obtain a first executable result, and send the first executable result to the verifier; The verifier is used to perform verification based on the first executable result and obtain a verification result.

2. The system according to claim 1, characterized in that, The planning server includes: a node filtering unit; The node filtering unit is used to filter tasks in the node description template according to the requirements of the first scenario to obtain the first node type.

3. The system according to claim 2, characterized in that, The planning server also includes: a component filtering unit; The component filtering unit is used to filter tasks according to the requirements of the first scenario to obtain the number of the first nodes.

4. The system according to claim 3, characterized in that, The component filtering unit includes: a first calling module; The first calling module is used to obtain a first calling tool according to the first scenario requirements, and to filter the first scenario requirements through the first calling tool to obtain the first number of nodes.

5. The system according to claim 4, characterized in that, The planning server also includes: a first reflection unit; The first reflection unit is used to reflect on the first node information, obtain a first reflection result, and determine whether to make adjustments based on the first reflection result.

6. The system according to claim 5, characterized in that, The orchestration server includes: a structural orchestration unit and an edge module; The structure arrangement unit is used to fill the first node information into the edge module to obtain the first parameter template.

7. The system according to claim 6, characterized in that, The structure orchestration unit includes: a second invocation unit; The second calling unit is used to obtain a second calling tool based on the first node information, and to fill the first node information into the edge module through the second calling tool.

8. The system according to claim 7, characterized in that, The structural arrangement unit also includes: a connection module; The connection module is used to obtain a first number of connections based on the first node information, and then perform node connection.

9. The system according to claim 8, characterized in that, The orchestration server also includes: a parameter filling unit and a node module; The parameter filling unit is used to concatenate the first parameter template with the node module to obtain a first executable result.

10. The system according to claim 9, characterized in that, The orchestration server further includes: a second reflection unit; The second reflection unit is used to reflect based on the first parameter template, obtain a second reflection result, and determine whether to make adjustments based on the second reflection result.

Images