System and method for converting natural language into model workflow based on AI Agent

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a system and a method for converting a natural language into a model workflow based on an AI Agent, and realizes automation of converting the natural language into the model workflow by defining a modeling analysis specific field language, a language analysis engine based on the AI Agent and a workflow generation engine. The AI Agent plays a core role in the AI Agent, the AI Agent has high intelligence and adaptability through strong natural language understanding, knowledge reasoning and scheduling control capabilities and high expansibility, and the natural language is converted into a model workflow for modeling analysis based on the AI Agent. The efficiency of interaction between the user and modeling analysis is improved, and the use threshold is lowered. By means of an AI Agent + large language model, a user can complete initial configuration of a modeling analysis model only by using a natural language.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a system and method for converting natural language into a model workflow based on an AI Agent. Background Art

[0002] AI Agent is a "system driven by a large language model as its brain" that has the ability to autonomously understand, perceive, plan, remember and use tools. It can automatically perform complex tasks and can also be called an artificial intelligence entity.

[0003] In the existing workflow technology, it generally includes two parts: process engine and process designer, which are the core components of workflow. The process designer is used to plan and design the workflow; the process engine provides analysis of the process and drives the business process to flow as designed. When dealing with the above-mentioned approval flow business, these two components are manually configured, and there is a problem that the flow nodes cannot be automatically identified, and the workflow cannot automatically adapt to business changes. This results in defects such as low work efficiency, high maintenance costs, and poor self-adaptation.

[0004] Existing technical solutions may be:

[0005] 1. Manually configure the model workflow: Manually configure the model workflow and node parameters according to the scenario requirements, but the model development threshold is high: users need to learn and accumulate the functions and skills of model workflow configuration, such as SQL syntax and Python syntax, which also increases the learning cost.

[0006] Low efficiency of human-computer interaction: When configuring model nodes, there are many similar operations. For example, common input and output can be abstracted into a unified processing flow, reducing the steps that require manual configuration and improving interaction efficiency.

[0007] 2. For example, the patent with publication number CN113052544A discloses a method, device and storage medium for intelligently adapting a workflow according to user behavior. The disclosed method is:

[0008] S1: Document entry: Use template library to standardize the entry method and template common terms and official standard terms; analyze and process the content of the document, extract key information and accumulate user behavior data;

[0009] S2: Intelligent workflow adaptation: Process the information entered into the approval document to obtain key information; conduct comprehensive analysis and integration of the information obtained to obtain flow node information and workflow adaptation parameters; and automatically adapt to new workflows and flow nodes according to the adaptation parameters;

[0010] S3: Workflow drive: After processing in step S2, the system drives the approval flow business to continue to flow downward through the newly adapted workflow.

[0011] 3. For example, the patent with publication number WO2022160707A1 discloses a human-computer interaction method, device, storage medium and electronic device combining RPA and AI, and its specific method is:

[0012] S1: Use Robotic Process Automation (RPA) to obtain the company’s business-related data and workflow data

[0013] S2: Using artificial intelligence AI to process the business-related data to obtain structured target business data corresponding to the business-related data

[0014] S3: Generate response processing logic corresponding to the enterprise based on the target business data and the workflow data in combination with the preconfigured rule engine

[0015] S4: Controlling the digital employee execution device to interact with the user according to the response processing logic.

[0016] As for the existing technologies introduced in 2 and 3, it is necessary to first learn the target business data and target workflow data, and generate the response processing logic corresponding to the enterprise in combination with the pre-configured rule engine; the pre-learning cost is high in the early stage, and new rules and response processing logic need to be generated according to different enterprise businesses. This solution is not suitable for general modeling and analysis processes. Summary of the invention

[0017] In view of the above problems, the present invention provides a system and method for converting natural language into a model workflow based on AI Agent.

[0018] The technical solution adopted is a system based on AI Agent to convert natural language into model workflow, including modeling and analysis of specific domain language modules, AI Agent-language parsing engine and AI Agent-workflow generation engine;

[0019] The modeling and analysis domain-specific language module provides data sources, data processing operations, and model types for the AI ​​Agent-language parsing engine and the AI ​​Agent-workflow generation engine;

[0020] The AI ​​Agent-language parsing engine is used to convert natural language instructions into standardized OpenWDL semantic data;

[0021] The AI ​​Agent-workflow generation engine is used to construct the generated OpenWDL semantic data into a model workflow.

[0022] Optionally, in the modeling and analysis domain-specific language module, the data source includes a resource directory, a database table, and a data API;

[0023] The data processing operations include reading, cleaning, transforming, aggregating, associating and filtering;

[0024] The model types include preprocessing, SQL calculations, and Python scripts.

[0025] Based on the above system, this application also provides a method for converting natural language into a model workflow based on AI Agent, including the following steps:

[0026] S1. In the AI ​​Agent-language parsing engine, convert natural language into template prompt words;

[0027] S2. In the AI ​​Agent-language parsing engine, the prompt word is converted into OpenWDL semantic data, and the generated OpenWDL semantic data is passed to the AI ​​Agent-workflow generation engine;

[0028] S3. In the AI ​​Agent-workflow generation engine, the OpenWDL semantic data is parsed;

[0029] S4. In the AI ​​Agent-workflow generation engine, the parsed data is combined with the model rule library to generate an abstract logic model workflow;

[0030] S5. In the AI ​​Agent-workflow generation engine, the abstract logic model workflow is converted into Agent task planning;

[0031] S6. Build the model workflow in the AI ​​Agent-Workflow Generation Engine.

[0032] Optionally, in S1, keywords of natural language instructions are extracted and annotated by modeling and analyzing data sources in a specific domain language module.

[0033] Optionally, in S1, add instruction-related model type examples to optimize output quality, and combine with built-in model templates to generate optimized prompt words.

[0034] Optionally, in S2, the prompt words are converted into OpenWDL semantic data through the language conversion function of the large language model.

[0035] Optionally, S3 includes the following sub-steps:

[0036] S31 extracts the input parameters in the OpenWDL semantic data, the call information covering the dependencies between the calls, and the output parameters corresponding to each call and a Task;

[0037] S32. Extract the input parameters, execution commands and output parameters of the Task.

[0038] Optionally, S4 includes the following sub-steps:

[0039] S41. Establishing logical relationships based on the model rule base;

[0040] S42. Following the model node rules, the Input, Command, and Output parameters of the Task are converted into the Model Input, Model Process, and Model Output parameters of the model node, respectively.

[0041] Optionally, S5 includes the following sub-steps:

[0042] S51. Define the objectives and constraints of the task;

[0043] S52. Build a rule base containing tasks, actions and conditions;

[0044] S53. Based on the task and rule base, the abstract logic model workflow is converted into Agent task planning.

[0045] Optionally, in S6, Action is automatically triggered to execute task planning, and the modeling API is called to build a model workflow and collect feedback on the execution results.

[0046] The benefits of the present invention include:

[0047] 1. By defining a modeling and analysis domain-specific language (DSL), the AI ​​Agent-based language parsing engine and workflow generation engine realize the automation of converting natural language into model workflow. AI Agent plays a core role in this process. Through its powerful natural language understanding, knowledge reasoning, scheduling control capabilities and high scalability, the intelligent agent has a high degree of intelligence and adaptability.

[0048] 2. Model workflow based on artificial intelligence to convert natural language into modeling and analysis. This improves the efficiency of user interaction with modeling and analysis and lowers the threshold for use. With the help of AI Agent + large language model, users only need to use natural language to complete the initial configuration of the modeling and analysis model.

[0049] 3. Based on the AI ​​Agent-language parsing engine, natural language is parsed into standardized OpenWDL semantic data. By adjusting the prompt words, it can adapt to various security scenario requirements, improve the efficiency of model incubation, and improve the efficiency of human-computer interaction. Users do not need to master complex programming languages, but only need to use DSL to describe the complex model building process, which lowers the threshold for model development. DSL shields the underlying technical details, allowing non-professionals to participate in model development and enhances the interpretability of the model.

[0050] 4. Based on the AI ​​Agent-workflow generation engine, the OpenWDL semantic data is generated into a modeling and analysis model workflow. The Agent actively calls the modeling and analysis platform API to generate the model workflow, automatically completing the modeling process, allowing users to focus on the business. The automated model workflow generation greatly reduces manual intervention and improves work efficiency. Intelligent workflow optimization continuously improves the quality of workflow generation by adding feedback mechanisms and self-correction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The figure is a schematic diagram of the Agent framework. DETAILED DESCRIPTION

[0052] The following describes the implementation of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific implementations, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0053] It should be noted that the illustrations provided in the following embodiments are only used to schematically illustrate the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0054] A system based on AI Agent that converts natural language into model workflow, including a modeling and analysis domain-specific language module, an AI Agent-language parsing engine, and an AI Agent-workflow generation engine;

[0055] The modeling and analysis domain-specific language module provides data sources, data processing operations, and model types for the AI ​​Agent-language parsing engine and the AI ​​Agent-workflow generation engine;

[0056] The AI ​​Agent-language parsing engine is used to convert natural language instructions into standardized OpenWDL semantic data;

[0057] The AI ​​Agent-workflow generation engine is used to construct the generated OpenWDL semantic data into a model workflow.

[0058] The purpose of this design is to achieve the automation of converting natural language into model workflow by defining a specific domain language for modeling and analysis, and based on the language parsing engine and workflow generation engine of AI Agent. AI Agent plays a core role in this, and through its powerful natural language understanding, knowledge reasoning, scheduling control capabilities and high scalability, the intelligent agent has a high degree of intelligence and adaptability.

[0059] Meanwhile, in this embodiment, in the modeling and analysis domain-specific language module, the data source includes a resource directory, a database table, and a data API;

[0060] The data processing operations include reading, cleaning, transforming, aggregating, associating and filtering;

[0061] The model types include preprocessing, SQL calculations, and Python scripts.

[0062] Based on the above system, this embodiment also provides a method for converting natural language into a model workflow based on an AI Agent, including the following steps:

[0063] S1. In the AI ​​Agent-language parsing engine, convert natural language into template prompt words;

[0064] S2. In the AI ​​Agent-language parsing engine, the prompt word is converted into OpenWDL semantic data, and the generated OpenWDL semantic data is passed to the AI ​​Agent-workflow generation engine;

[0065] S3. In the AI ​​Agent-workflow generation engine, the OpenWDL semantic data is parsed;

[0066] S4. In the AI ​​Agent-workflow generation engine, the parsed data is combined with the model rule library to generate an abstract logic model workflow;

[0067] S5. In the AI ​​Agent-workflow generation engine, the abstract logic model workflow is converted into Agent task planning;

[0068] S6. Build the model workflow in the AI ​​Agent-Workflow Generation Engine.

[0069] Among them, in S1, the keywords of natural language instructions are extracted and labeled by modeling and analyzing the data source in the specific domain language module.

[0070] At the same time, in S1, examples of model types related to instructions are added to optimize the output quality, and the built-in model templates are combined to generate optimized prompt words.

[0071] Furthermore, in S2, the prompt words are converted into OpenWDL semantic data through the language conversion function of the large language model.

[0072] Furthermore, S3 includes the following sub-steps:

[0073] S31 extracts the input parameters in the OpenWDL semantic data, the call information covering the dependencies between the calls, and the output parameters corresponding to each call and a Task;

[0074] S32. Extract the input parameters, execution commands and output parameters of the Task.

[0075] Meanwhile, S4 includes the following sub-steps:

[0076] S41. Establishing logical relationships based on the model rule base;

[0077] S42. Following the model node rules, the Input, Command, and Output parameters of the Task are converted into the Model Input, Model Process, and Model Output parameters of the model node, respectively.

[0078] Furthermore, S5 includes the following sub-steps:

[0079] S51. Define the objectives and constraints of the task;

[0080] S52. Build a rule base containing tasks, actions and conditions;

[0081] S53. Based on the task and rule base, the abstract logic model workflow is converted into Agent task planning.

[0082] Finally, in S6, the Action is automatically triggered to execute the task planning, and the API is called to build the model workflow and collect feedback on the execution results.

[0083] The purpose of this design is to convert natural language into model workflows for modeling and analysis based on artificial intelligence. This improves the efficiency of user interaction with modeling and analysis and lowers the threshold for use. With the help of AI Agent + large language model, users only need to use natural language to complete the initial configuration of the modeling and analysis model.

[0084] Based on the AI ​​Agent-language parsing engine, natural language is parsed into standardized OpenWDL semantic data. By adjusting the prompt words, it can adapt to various security scenario requirements, improve the efficiency of model incubation, and improve the efficiency of human-computer interaction. Users do not need to master complex programming languages, but only need to use DSL to describe the complex model building process, which lowers the threshold for model development. DSL shields the underlying technical details, allowing non-professionals to participate in model development, enhancing the interpretability of the model.

[0085] Based on the AI ​​Agent-workflow generation engine, OpenWDL semantic data is generated into a modeling and analysis model workflow. The Agent actively calls the modeling and analysis platform API to generate the model workflow and automatically completes the modeling process, allowing users to focus on their business. Automated model workflow generation greatly reduces manual intervention and improves work efficiency. Intelligent workflow optimization continuously improves the quality of workflow generation by adding feedback mechanisms and self-correction capabilities.

[0086] In a specific example, take the generation of an IP asset activity analysis model as an example:

[0087] Preprocessing: Read data from the original database TCP log, extract IP and time fields, clean up missing IP values, filter out data with communication load packet greater than 0, and obtain the tcp_tmp table;

[0088] Then SQL calculation: the service IP of the tcp_tmp table data is associated with the IP of the filter configuration library asset table, and the IP access frequency is calculated according to the trigger time dimension to obtain an IP asset activity table;

[0089] The final Python script: labels the IP asset activity table: geographic location, source access frequency, and generates an IP asset activity table.

[0090] At the same time, in S1, keyword extraction: TCP log, IP asset table, trigger time, IP access frequency, source access frequency, labeling, IP asset activity table

[0091] Annotation:

[0092] -TCP Log: Input Table

[0093] -IP Asset Table: Input Table

[0094] -Trigger time: Analysis dimension, dimension statistics according to trigger time

[0095] -IP access frequency: analysis indicator

[0096] -Source access frequency: analysis indicator

[0097] -Tag: trigger tag task

[0098] -IP asset activity table: Output table

[0099] After the above processing, the optimized prompt word "Generate an analysis model based on TCP logs to calculate the activity of IP assets" is obtained.

[0100] Data preprocessing:

[0101] Input table: TCP log table

[0102] Processing process: clean the original log data, extract the client_ip, server_ip, trigger time (date), communication load (packet_payload) and other field information, filter packet_payload>0, and process missing values ​​and abnormal values

[0103] Output table: Temporary output table: tcp_tmp_log

[0104] SQL compute node:

[0105] Input tables: tcp_tmp_log table, IP asset table

[0106] Processing process: tcp_tmp_log table server_ip is associated with the query IP asset table, and the trigger time is used as the dimension, the time granularity is minutes, and the server_ip access frequency, time distribution and other indicators are calculated.

[0107] Output table: ip asset access record table (ip_asset_histroy_log)

[0108] Python script node

[0109] Input table: ip_asset_histroy_log table

[0110] Processing: Tag the IP asset access record table, parse the geographic location, longitude and latitude based on client_ip, calculate and mark the access frequency (high, medium, low)

[0111] Output table: ip asset activity table (ip_asset_active_log)

[0112] Output:

[0113] Convert the above process description into a workflow that complies with the OpenWDL specification.

[0114] In step S2, the WorkFlow example generated based on the prompt word:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] In step S5, the abstract logic model workflow is converted into Agent task planning.

[0123] Define the task: clearly define the objectives and constraints of the task. (Complete the creation of the model, the addition of model nodes, the configuration of node parameters, and the setting of the relationship between nodes in sequence, thus generating a comprehensive and complete model workflow)

[0124] Define a rule base: Build a rule base that includes tasks, actions, and conditions. (For example: convert the workflow Task into a Model Node, call the modeling API interface, and get a correct response before proceeding to the next step.)

[0125] Generate task planning: Based on the tasks and rule base, convert the abstract logic model workflow into agent task planning.

[0126] Finally, in step S6, the task planning is executed: the corresponding actions are executed according to the generated task plan. Feedback mechanism: feedback on the execution results is collected to update the model and optimize future planning.

[0127] like Figure 1 As shown, the Agent framework is described:

[0128] Mission Planning

[0129] Task splitting: Complex tasks cannot be solved in one go and need to be split into multiple parallel or serial subtasks for solution. The goal of task planning is to find an optimal route that can solve the problem.

[0130] Self-reflection: Self-reflection is an important aspect that allows autonomous agents to iteratively improve by refining past action decisions and correcting previous mistakes. It plays a vital role in real-world tasks where trial and error are inevitable. ReAct (Yao et al. 2023) found that when asking agents to perform the next action, adding LLM's own thinking process, and putting the thinking process, execution tools and parameters, and execution results in the prompt, the model can have a better ability to reflect on the current and previous task completion, thereby improving the model's problem-solving ability.

[0131] Thought:...

[0132] Action:...

[0133] Observation:...

[0134] ...(repeat the above process)

[0135] Chain of Thoughts: has become a standard prompting technique to enhance model performance on complex tasks. The model is instructed to "think one step at a time" to leverage more test time computation by breaking down difficult tasks into smaller, simpler steps. CoT turns large tasks into multiple manageable ones and illuminates the explanation of the model's thought process.

[0136] Thinking Tree: Expands CoT by exploring multiple reasoning possibilities at each step. It first breaks down the problem into multiple thinking steps and generates multiple thinking in each step, creating a tree structure. The search process can be BFS (breadth first search) or DFS (depth first search), and each state is evaluated by a classifier (through a hint) or majority voting.

[0137] Memory

[0138] Perceptual memory: This is the earliest stage of memory and provides the ability to retain an impression of sensory information (visual, auditory, etc.) after the original stimulus has ended. Perceptual memory typically only lasts a few seconds. Subcategories include iconic memory (vision), echoic memory (auditory), and tactile memory (touch). Perceptual memory serves as a learned embedded representation of the original input, which can be text, images, or other modalities.

[0139] Short-term memory: It stores information of which we are currently aware and which is needed to perform complex cognitive tasks such as learning and reasoning. Short-term memory is thought to have a capacity of about 7 items (Miller 1956) and lasts for 20-30 seconds. Short-term memory serves as contextual learning. It is short and limited because it is constrained by the length of the Transformer's finite context window.

[0140] Long-term memory (LTM): Long-term memory can store information for a considerable period of time, ranging from a few days to decades, and the storage capacity is essentially unlimited. There are two subtypes of LTM:

[0141] 1. Explicit / declarative memory: This is the memory for facts and events. It refers to those memories that can be consciously recalled and includes episodic memory (events and experiences) and semantic memory (facts and concepts).

[0142] 2. Implicit / procedural memory: This type of memory is unconscious and involves skills and routines that are performed automatically, such as riding a bicycle or typing on a keyboard.

[0143] Long-term memory acts as a storage of external vectors that the agent can process when querying, accessible via fast retrieval.

[0144] Tools

[0145] Agent learning calls external APIs to obtain additional information missing from the model weights (which are often difficult to change after pre-training), including current information, code execution capabilities, access to proprietary information sources, etc.

[0146] API-Bank (Li et al. 2023) is a benchmark for evaluating the performance of tool-enhanced LLM. It contains 53 commonly used API tools, a complete tool-enhanced LLM workflow, and 264 annotated conversations involving 568 API calls. The selection of APIs is very diverse, including search engines, calculators, calendar queries, smart home control, schedule management, health data management, account authentication workflows, etc. Because there are a large number of APIs, LLM can first access the API search engine to find suitable API calls, and then use the corresponding documentation to make calls.

[0147] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A system based on AI Agent that converts natural language into model workflow, characterized by: Includes modeling and analysis of specific domain language modules, AI Agent-language parsing engine and AI Agent-workflow generation engine; The modeling and analysis domain-specific language module provides data sources, data processing operations and model types for the AI ​​Agent-language parsing engine and the AI ​​Agent-workflow generation engine; The AI ​​Agent-language parsing engine is used to convert natural language instructions into standardized OpenWDL semantic data; The AI ​​Agent-workflow generation engine is used to construct the generated OpenWDL semantic data into a model workflow.

2. The system for converting natural language into model workflow based on AI Agent according to claim 1, characterized in that: In the modeling and analysis domain-specific language module, data sources include resource directories, database tables, and data APIs; The data processing operations include reading, cleaning, transforming, aggregating, associating and filtering; The model types include preprocessing, SQL calculations, and Python scripts.

3. A method for converting natural language into model workflow based on AI Agent, based on the system for converting natural language into model workflow based on AI Agent in claim 2, characterized in that: The following steps are involved: S1. In the AI ​​Agent-language parsing engine, convert natural language into template prompt words; S2. In the AI ​​Agent-language parsing engine, the prompt word is converted into OpenWDL semantic data, and the generated OpenWDL semantic data is passed to the AI ​​Agent-workflow generation engine; S3. In the AI ​​Agent-workflow generation engine, the OpenWDL semantic data is parsed; S4. In the AI ​​Agent-workflow generation engine, the parsed data is combined with the model rule library to generate an abstract logic model workflow; S5. In the AI ​​Agent-workflow generation engine, the abstract logic model workflow is converted into Agent task planning; S6. Build the model workflow in the AI ​​Agent-Workflow Generation Engine.

4. The method for converting natural language into model workflow based on AI Agent according to claim 3, characterized in that: In S1, keywords of natural language instructions are extracted and annotated by modeling and analyzing the data source in the domain-specific language module.

5. The method for converting natural language into model workflow based on AI Agent according to claim 4, characterized in that: In S1, examples of model types related to instructions are added to optimize output quality, and the built-in model templates are combined to generate optimized prompt words.

6. The method for converting natural language into model workflow based on AI Agent according to claim 3, characterized in that: In S2, the prompt words are converted into OpenWDL semantic data through the language conversion function of the large language model.

7. The method for converting natural language into model workflow based on AI Agent according to claim 3, characterized in that: S3 includes the following sub-steps: S31 extracts the input parameters in the OpenWDL semantic data, the call information covering the dependencies between the calls, and the output parameters corresponding to each call and a Task; S32. Extract the input parameters, execution commands and output parameters of the Task.

8. The method for converting natural language into model workflow based on AI Agent according to claim 3, characterized in that: S4 includes the following sub-steps: S41. Establishing logical relationships based on the model rule base; S42. Follow the model node rules and convert the Input, Command, and Output parameters of the Task into the Model Input, Model Process, and Model Output parameters of the model node, respectively.

9. The method for converting natural language into model workflow based on AI Agent according to claim 8, characterized in that: S5 includes the following sub-steps: S51. Define the objectives and constraints of the task; S52. Build a rule base containing tasks, actions and conditions; S53. Based on the task and rule base, the abstract logic model workflow is converted into Agent task planning.

10. The method for converting natural language into model workflow based on AIAgent according to claim 9, characterized in that: In S6, Action is automatically triggered to execute task planning, and the modeling API is called to build the model workflow and collect feedback on the execution results.

Images