An RPA command automatic generation system and method based on AI Agent

Abstract

The present invention discloses an RPA command automatic generation system and method based on AI Agent, which relates to the field of process automation. The system includes: taking the perception unit, planning unit, memory unit, action unit and tool use unit of the AI ​​Agent system as the basic architecture, constructing an RPA command automatic generation system including a user demand module, a planning module, a process plan knowledge base, an execution module, a service call module, a process service knowledge base, an automated process command module and a process command service module; the user demand module belongs to the perception unit, the planning module belongs to the planning unit; the process plan knowledge base and the process service knowledge base belong to the memory unit; the execution module, the service call module and the automated process command module belong to the action unit; and the process command service module belongs to the tool use unit. The present invention enhances the ability to automatically generate process commands.

Description

Technical Field

[0001] The present invention relates to the field of process automation, and in particular to an AI Agent-based RPA command automatic generation system and method. Background Art

[0002] Robotic Process Automation (RPA) is also known as a software robot. With RPA technology, you can configure RPA scripts to allow robots to simulate human user operations, interact with various software in the operating system, and perform repetitive or error-prone business process tasks, such as processing transactions and replying to emails. RPA technology can quickly and accurately handle a large number of repetitive process tasks, improve the efficiency of process execution, and reduce the probability of errors during process execution. However, with the increase in process commands and the complexity of the process, the writing of RPA scripts has become very cumbersome. Configuring a high-quality RPA script will consume a lot of human resources, which has hindered the application and implementation of RPA technology.

[0003] AI Agent can also be called "intelligent agent" or "intelligent business assistant", specifically referring to an intelligent agent based on a large language model. Its cooperation and task completion capabilities mainly come from the powerful semantic understanding and decision-making capabilities of the large language model behind it, allowing people to perform and handle professional or complex work tasks in a highly automated manner using natural language as an interactive method, thereby releasing personnel energy to a great extent. LLM is used as the main component of the brain or controller of AI Agent.

[0004] Large Language Model (LLM) has demonstrated outstanding performance in natural language text generation and code generation tasks due to its powerful semantic understanding capabilities on natural corpus. However, in the field of robotic process automation, the numerous process commands provide various paths for the completion of process tasks. At the same time, each process command has a unique configuration syntax, which poses a challenge to the automatic generation of process commands. Therefore, relying solely on large language models is not enough for the task of automatic generation of process commands. An adaptive architecture is needed to apply the powerful semantic understanding capabilities of large language models to the field of robotic process automation and promote the wider application and implementation of RPA technology.

[0005] Traditional RPA can only process work under given conditions according to the preset processes in the program, while AI Agent can interact with the environment, perceive information and make corresponding thoughts and actions. Summary of the invention

[0006] The purpose of this invention is to solve the problem that traditional RPA process commands are based on preset rather than automatic generation, and propose an RPA command automatic generation system based on AI Agent, which can interact with the environment through AI Agent, perceive information and make corresponding thoughts and actions, including:

[0007] Based on the AI ​​Agent system's perception unit, planning unit, memory unit, action unit, and tool usage unit, we build an RPA command automatic generation system that includes a user demand module, a planning module, a process planning knowledge base, an execution module, a service call module, a process service knowledge base, an automated process command module, and a process command service module.

[0008] Among them, the user demand module belongs to the perception unit, the planning module belongs to the planning unit; the process planning knowledge base and the process service knowledge base belong to the memory unit; the execution module, the service call module and the automated process command module belong to the action unit; the process command service module belongs to the tool use unit;

[0009] The specific functions of each module are as follows:

[0010] The user demand module is used to obtain user needs;

[0011] The planning module understands user needs and combines the process planning knowledge in the process planning knowledge base to generate a process plan that can meet user needs;

[0012] The process planning knowledge base is used to store process planning knowledge;

[0013] The execution module understands the process plan and generates the next service call intention based on the current process execution environment status. If the process plan is executable, the intention is passed to the service call module; if the process plan is not executable, the information that the process plan is not executable is fed back to the planning module;

[0014] The service call module understands the current service call intention, refers to the process service knowledge base, and generates a call to the corresponding process command service;

[0015] The process command service module receives the call of the process command service generated by the service call module, generates a process command or reports an error;

[0016] When the service call module receives an error from the process command service module, it regenerates the call of the corresponding process command service, or feeds back to the execution module, and the execution module regenerates the service call intent;

[0017] After the service calling module receives the process command from the process command service module, the service calling module records the command generated by the process command service module;

[0018] When the execution module determines that the task is completed, it outputs the complete process command recorded by the service call module to the automation process command module.

[0019] Furthermore, the planning module understands user needs specifically as follows: the planning module uses the semantic understanding ability of the large language model to extract the semantics of user needs, perform reasoning, and generate reasoning results.

[0020] Furthermore, the planning module understands the user's needs and combines the process planning knowledge in the process planning knowledge base to generate a process plan that can meet the user's needs, specifically:

[0021] According to the user demand D understood by the planning module, the similarity between the user demand D and the vector of the process planning knowledge in the process planning knowledge base is used to find the fragment with the highest similarity to demand D in the process planning knowledge base using the vector search method to obtain the process planning knowledge related to the user demand. ;

[0022] Combine user needs D and process planning knowledge , using the semantic understanding ability of the large language model to decompose user needs into several sub- , form a plan .

[0023] Furthermore, the execution module understands the process plan, combines the current process execution environment status, generates the next service call intention, and if the process plan is executable, passes the intention to the service call module; if the process plan is not executable, the information that the process plan is not executable is fed back to the planning module;

[0024] Specifically:

[0025] The execution module receives the process plan P output by the planning module, observes the state S from the environment, inputs the process plan P and the environment state S into the large language model, and the large language model outputs the service call intention a. ;

[0026] If the large language model believes that plan P can continue to be executed under the current environment state S, the service call intention a is passed to the service call module, and the service call module calls a specific service according to the service call intention;

[0027] If the large language model believes that plan P cannot handle the problem presented in the current environment state S, it will pass the relevant error information back to the planning module, and the planning module will regenerate the plan based on the error information and specific information of the environment.

[0028] Furthermore, the service call module understands the current service call intention, refers to the process service knowledge base, and generates a call to the corresponding process command service, specifically:

[0029] The service call module receives the current service call intention a and uses vector search technology to search the process service knowledge base Find the service information that matches the service call intent a , ;

[0030] Large language model reference service information The service description and format description in the above code are used to generate a call to the service with the goal of satisfying the service call intention a, and the call process command service module generates a robot process automation command for the corresponding service.

[0031] Furthermore, when the service calling module receives an error from the process command service module, it regenerates the call of the corresponding process command service, or feeds back to the execution module, and the execution module regenerates the service calling intent, specifically:

[0032] When the process command service module reports an error, the service call module makes a judgment. If the call format is wrong, the service call module regenerates the call format; if it is because the current service call intention a cannot meet the requirements, it is fed back to the execution module, and the execution module re-outputs the service call intention.

[0033] Furthermore, the service calling module records the command generated by the process command service module into the context environment in which the command is generated.

[0034] The present invention also proposes an automatic generation method of RPA commands for AI Agent, which is based on the above system implementation and includes:

[0035] Based on the AI ​​Agent system's perception unit, planning unit, memory unit, action unit, and tool usage unit, we build an RPA command automatic generation system that includes a user demand module, a planning module, a process planning knowledge base, an execution module, a service call module, a process service knowledge base, an automated process command module, and a process command service module.

[0036] Among them, the user demand module belongs to the perception unit, the planning module belongs to the planning unit; the process planning knowledge base and the process service knowledge base belong to the memory unit; the execution module, the service call module and the automated process command module belong to the action unit; the process command service module belongs to the tool use unit;

[0037] Use the user requirements module to obtain user requirements;

[0038] Use the planning module to understand user needs and combine the process planning knowledge in the process planning knowledge base to generate a process plan that can meet user needs;

[0039] Use process planning knowledge base to store process planning knowledge;

[0040] Use the execution module to understand the process plan, combine it with the current process execution environment status, generate the next service call intention, if the process plan is executable, pass the intention to the service call module; if the process plan is not executable, feed back the information that the process plan is not executable to the planning module;

[0041] Use the service call module to understand the current service call intention, refer to the process service knowledge base, and generate the call of the corresponding process command service;

[0042] Use the process command service module to receive the call of the process command service generated by the service call module, generate a process command or report an error;

[0043] When the service call module receives an error from the process command service module, it regenerates the call of the corresponding process command service, or feeds back to the execution module, and the execution module regenerates the service call intent;

[0044] When the service calling module receives the process command from the process command service module, the service calling module records the command generated by the process command service module;

[0045] When the execution module determines that the task is completed, it outputs the complete process command recorded by the service call module to the automation process command module.

[0046] The beneficial effects brought by the technical solution provided by the present invention are:

[0047] The present invention proposes an automatic generation system and method of robot process automation commands for AI Agent, which is based on the brain module, perception module and action module of the AI ​​Agent system to construct an RPA command automatic generation system including: user demand module, planning module, process planning knowledge base, execution module, service call module, process service knowledge base, automated process command module and process command service module; the execution module understands the process plan, generates service call intentions in combination with the current environment state; the service call module understands the current service call intentions, refers to the process service knowledge base, and generates calls to the corresponding process command services; through communication between modules and calls to services, a complete robot process automation command is finally generated. The planning module, execution module, service call module and automated process command module communicate with each other, which enhances the ability to automatically generate process commands. By introducing the process planning knowledge base and the process service knowledge base, expert knowledge in the field of robot process automation is introduced into the large language model, which reduces the illusion phenomenon and improves the reasoning ability of the large language model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1It is a block diagram of an AI Agent-based RPA command automatic generation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0050] Example 1: A block diagram of an AI Agent-based RPA command automatic generation system according to an embodiment of the present invention is shown in FIG. Figure 1 , specifically including:

[0051] Based on the perception unit, planning unit, memory unit, action unit, and tool usage unit of the AI ​​Agent system, an RPA command automatic generation system is constructed, which includes a user demand module, a planning module, a process planning knowledge base, an execution module, a service call module, a process service knowledge base, an automated process command module, and a process command service module.

[0052] The perception unit is the interface for the agent to interact with the external environment and is responsible for collecting and parsing environmental data. Planning is the decision-making center of the agent, which breaks down the goal into executable steps and formulates strategies to achieve the goal. The memory unit allows the agent to store and retrieve information, supports learning and long-term knowledge accumulation, and provides the agent with information storage and retrieval capabilities. The tool use unit is the process by which the agent uses external resources or tools to enhance its perception, decision-making and action capabilities. These tools can be APIs, software libraries, hardware devices or other services. The action unit is the specific implementer of the agent to perform tasks and interact with the environment. The specific actions performed based on planning and memory are the agent's response to the environment and the completion of tasks.

[0053] Among them, the user demand module belongs to the perception unit, the planning module belongs to the planning unit; the process planning knowledge base and the process service knowledge base belong to the memory unit; the execution module, the service call module and the automated process command module belong to the action unit; and the process command service module belongs to the tool usage unit.

[0054] In the preferred embodiment of the present invention, the specific functions of each module are as follows:

[0055] The user demand module is used to obtain user demand D and copy the content in the document origin.docx to the document res.docx.

[0056] The planning module uses the semantic understanding ability of the large language model to extract the semantics of user needs, perform reasoning, and generate reasoning results. The planning module understands the user demand D based on the reasoning results, combines the process planning knowledge in the process planning knowledge base, and uses the similarity between the vectors of user demand D and the process planning knowledge in the process planning knowledge base. The vector search method is used to find the fragment with the highest similarity to demand D in the process planning knowledge base, and obtain the process planning knowledge related to the user demand. Specifically, the planning module receives user demand D and uses the retrieval-augmented generation (RAG) technology to generate the process planning knowledge base through pre-defined prompt words. Find process planning knowledge related to user demand D ,Right now . Combine user needs D and process planning knowledge , using the semantic understanding ability of the large language model to decompose user needs into several sub- , forming a process plan ,Right now . And send the process plan to the execution module. Among them, the process plan knowledge base is used to store the process plan knowledge.

[0057] The execution module understands the process plan, combines the current environment status of the process execution, and generates the next service call intention. If the process plan is executable, the intention is passed to the service call module; if the process plan is not executable, the information that the process plan is not executable is fed back to the planning module. Specifically, the execution module receives the process plan P output by the planning module, observes the state S from the environment, inputs the process plan P and the environment state S into the large language model, and the large language model outputs the service call intention a (i.e., action a). ; If the large language model believes that plan P can continue to be executed under the current environment state S, the service call intention a is passed to the service call module, and the service call module calls the specific service according to the service call intention; if the large language model believes that plan P is not sufficient to handle the problems presented in the current environment state S, the information that plan p is not executable is passed to the planning module, and the planning module regenerates the plan based on the specific information of the environment.

[0058] The service call module understands the current service call intention. The service call module uses the semantic understanding ability of the large language model to extract the semantics of the service call intention, perform reasoning, and generate reasoning results. Refer to the process service knowledge base to generate the call of the corresponding process command service. Specifically, the service call module receives the current service call intention a and uses vector search technology to search the process service knowledge base. Find the service information that matches the service call intent a , Large Language Model Reference Service Information The service description and format description in the above code are used to generate a call to the service with the goal of satisfying the service call intention a, and the call process command service module generates a robot process automation command (process command) for the corresponding service.

[0059] The process command service module receives the call of the process command service generated by the service call module, generates a process command or reports an error.

[0060] When the service call module receives an error from the process command service module, it regenerates the call of the corresponding process command service, or feeds back to the execution module, and the execution module regenerates the service call intent. Specifically, when the process command service reports an error, the service call module makes a judgment. If the call format is wrong, the service call module regenerates the call format; if it is because the current service call intent a cannot meet the requirements, the information that the service call intent a cannot meet the requirements is fed back to the execution module, and the execution module re-outputs the service call intent.

[0061] After the service calling module receives the correct return information from the process command service module, the service calling module records the command generated by the process command service module, and the service calling module records the command generated by the process command service module into the context environment of command generation.

[0062] When the execution module determines that the task is completed, it outputs the complete process command stored in the context environment of the service call module to the automation process command module.

[0063] Embodiment 2: The present invention also proposes an automatic generation method of RPA commands for AI Agent, which is implemented based on the above system and includes:

[0064] Based on the AI ​​Agent system's perception unit, planning unit, memory unit, action unit, and tool usage unit, we build an RPA command automatic generation system that includes a user demand module, a planning module, a process planning knowledge base, an execution module, a service call module, a process service knowledge base, an automated process command module, and a process command service module.

[0065] Among them, the user demand module belongs to the perception unit, the planning module belongs to the planning unit; the process planning knowledge base and the process service knowledge base belong to the memory unit; the execution module, the service call module and the automated process command module belong to the action unit; the process command service module belongs to the tool use unit;

[0066] Use the user requirements module to obtain user requirements;

[0067] Use the planning module to understand user needs and combine the process planning knowledge in the process planning knowledge base to generate a process plan that can meet user needs;

[0068] Use process planning knowledge base to store process planning knowledge;

[0069] Use the execution module to understand the process plan, combine it with the current process execution environment status, generate the next service call intention, if the process plan is executable, pass the intention to the service call module; if the process plan is not executable, feed back the information that the process plan is not executable to the planning module;

[0070] Use the service call module to understand the current service call intention, refer to the process service knowledge base, and generate the call of the corresponding process command service;

[0071] Use the process command service module to receive the call of the process command service generated by the service call module, generate a process command or report an error;

[0072] When the service call module receives an error from the process command service module, it regenerates the call of the corresponding process command service, or feeds back to the execution module, and the execution module regenerates the service call intent;

[0073] When the service calling module receives the process command from the process command service module, the service calling module records the command generated by the process command service module;

[0074] When the execution module determines that the task is completed, it outputs the complete process command recorded by the service call module to the automation process command module.

[0075] Specifically, in the embodiment of the present invention, the implementation process of the actual application is:

[0076] Generate a plan:

[0077] (1) Open the document origin.docx; (2) Open the document res.docx; (3) Select all the contents in the document origin.docx; (4) Copy the contents in the document origin.docx; (5) Paste the copied contents into res.docx; (6) Save the document res.docx; (7) Close the document origin.docx; (8) Close the document res.docx.

[0078] The execution module understands the process plan. The execution module uses the semantic understanding ability of the large language model to extract the semantics of the process plan, perform reasoning, and generate reasoning results. The execution module receives the process plan P output by the planning module, observes the state S from the environment, inputs the process plan P and the environment state S into the large language model, and the large language model outputs the service call intention a. If the large language model believes that plan P can continue to be executed under the current environment state S, the service call intention a is passed to the service call module, and the service call module calls the specific service according to the received service call intention; if the large language model believes that plan P is not sufficient to handle the problems presented in the current environment state S (plan P is not executable), the information that plan P is not executable is passed to the planning module, and the planning module improves the plan according to the specific information of the environment.

[0079] Make an action decision: Currently, you need to open the document origin.docx first and send the decision to the service call module.

[0080] The service call module receives the service call intent, searches for services related to "open document origin.docx" in the process service knowledge base, and generates a service call WordOpenDocument("origin.docx") based on the description and call format of the related services.

[0081] After the WordOpenDocument service is called, the corresponding robot process automation command is generated according to the personalized configuration of the service provider, and the generated command is returned to the service calling module.

[0082] After receiving the return of the WordOpenDocument service, the service calling module determines that the service calling intention is satisfied, and then feeds back the successful result to the execution module, guiding the execution module to output the next service calling intention.

[0083] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AI Agent-based RPA command automatic generation system, characterized in that: include: Based on the perception, planning, memory, action, and tool usage units of the AI ​​Agent system, an RPA command automatic generation system is built, including several modules such as user needs, planning, execution, service call, automated process commands, process command services, as well as process planning and process service knowledge bases; The user demand module is a perception unit used to obtain user needs; The planning module is a planning unit used to understand user needs and generate process plans in combination with the process planning knowledge base; The process plan and process service knowledge base are memory units; The execution, service call and automated process command modules are action units. The execution module generates a service call intention based on the process plan. If the process plan is executable, the service call intention is passed to the service call module. If it is not executable, it is fed back to the planning module. The service call module generates a call to the process command service based on the service call intention and the process service knowledge base. Specifically: The execution module receives the process plan P output by the planning module, observes the state S from the environment, inputs the process plan P and the environment state S into the large language model, and the large language model outputs the service call intention a. ; If the large language model believes that plan P can continue to be executed under the current environment state S, the service call intention a is passed to the service call module, and the service call module calls a specific service according to the service call intention; If the large language model believes that plan P cannot handle the problem presented in the current environment state S, it will send the relevant error information back to the planning module, and the planning module will regenerate the plan based on the error information and specific information of the environment; The service call module receives the current service call intention a and uses vector search technology to search the process service knowledge base Find the service information that matches the service call intent a , ; Large language model reference service information The service description and format description in the above are used to generate a call to the service with the goal of satisfying the service call intention a, and the call process command service module generates a robot process automation command for the corresponding service; The process command service module is a tool usage unit, which generates process commands or reports errors based on the call of the process command service; the service call module receives the error and regenerates the call of the process command service, or feeds back to the execution module, and the execution module regenerates the service call intention; The service call module records the process commands and outputs the process commands to the automated process command module.

2. The AI ​​Agent-based RPA command automatic generation system according to claim 1, characterized in that: The planning module understands user needs specifically as follows: the planning module uses the semantic understanding ability of the large language model to extract the semantics of user needs, perform reasoning, and generate reasoning results.

3. The AI ​​Agent-based RPA command automatic generation system according to claim 1, characterized in that: The planning module understands user needs and generates a process plan based on the process plan knowledge base, specifically: According to the user demand D understood by the planning module, the similarity between the user demand D and the vector of the process planning knowledge in the process planning knowledge base is used to find the fragment with the highest similarity to demand D in the process planning knowledge base using the vector search method to obtain the process planning knowledge related to the user demand. ; Combine user needs D and process planning knowledge , using the semantic understanding ability of the large language model to decompose user needs into several sub- , form a plan .

4. The AI ​​Agent-based RPA command automatic generation system according to claim 1, characterized in that: The service call module receives an error and regenerates the call of the process command service, or feeds back to the execution module, which regenerates the service call intent, specifically: When the process command service module reports an error, the service call module makes a judgment. If the call format is wrong, the service call module regenerates the call format; if it is because the current service call intention a cannot meet the requirements, it is fed back to the execution module, and the execution module re-outputs the service call intention.

5. The AI ​​Agent-based RPA command automatic generation system according to claim 1, characterized in that: The service call module records the command generated by the process command service module into the context environment in which the command was generated.

6. A method for automatically generating RPA commands based on AI Agent, implemented based on the system according to any one of claims 1 to 5, characterized in that: include: Based on the perception, planning, memory, action, and tool usage units of the AI ​​Agent system, an RPA command automatic generation system is built, including several modules such as user needs, planning, execution, service call, automated process commands, process command services, as well as process planning and process service knowledge bases; The user demand module belongs to the perception unit, and the planning module belongs to the planning unit; The process planning knowledge base and process service knowledge base belong to the memory unit; the execution module, service call module and automated process command module belong to the action unit; The process command service module belongs to the tool usage unit; Use the user requirements module to obtain user requirements; Use the planning module to understand user needs and generate process plans in combination with the process planning knowledge base; Use the execution module to generate a service call intention based on the process plan. If the process plan is executable, the service call intention is passed to the service call module. If the process plan is not executable, it is fed back to the planning module. Use the service call module to generate a call for the process command service based on the service call intent and the process service knowledge base; Use the process command service module to generate process commands or report errors based on the call of the process command service; The service call module receives the error and regenerates the call of the process command service, or feeds back to the execution module, and the execution module regenerates the service call intent; Use the service call module to record process commands and output process commands to the automated process command module.

Images