Development system for generating agent from AI model
By clearly defining AI roles and data selection, and combining hybrid training and multi-dimensional verification, the problems of functional ambiguity and data redundancy in AI agent development have been solved, enabling efficient and targeted agent development.
Patent Information
- Application Number
- CN202511743046.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-17
AI Technical Summary
In the development of existing AI agents, unclear functional positioning and data redundancy lead to high training costs and low efficiency, as well as untimely verification feedback, which affect development efficiency and quality.
The AI role setting unit clarifies the functional positioning, the AI learning direction setting unit filters pure data, and the training process is optimized by using hybrid training algorithms and multi-dimensional verification to output efficient AI agents.
It enables efficient iteration and rapid deployment of AI agents, with highly targeted and adaptable functions, reducing development barriers and costs.
Smart Images
Figure CN121541858A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, and in particular to a development system for generating an agent from an AI model. BACKGROUND
[0002] With the rapid development of artificial intelligence technology, AI agents have been widely applied in various industries and fields. Different application scenarios have significantly different functional requirements for AI agents. For example, some scenarios require AI agents to focus on text classification functions, while some scenarios require AI agents to focus on recommendation functions or decision support functions.
[0003] However, there are many problems in the existing AI agent development process. On the one hand, the functional positioning of AI agents is not clear enough, and there is a lack of systematic setting of roles, goals, work processes and constraint conditions, which leads to ambiguous development direction of agents, weakly targeted functions, and difficulty in accurately meeting specific application requirements. On the other hand, the database relied on by AI agent training often contains a large amount of irrelevant data to the target function, and data redundancy not only increases the training cost, but also may interfere with the learning effect of the agent, resulting in low training efficiency and insufficient function accuracy of the agent.
[0004] In addition, in the existing agent development process, the training, verification and optimization links lack effective cooperation, the feedback of the verification result to the iteration of the agent is not timely and accurate enough, which further affects the development efficiency and quality of the AI agent. Therefore, there is an urgent need for an AI agent development system that can clearly define the function, ensure data purity, and achieve efficient iteration. SUMMARY
[0005] The purpose of the present application is to provide a development system for generating an agent from an AI model, to solve the problems of weakly targeted functions, redundant learning data and low development efficiency of AI agents in the prior art.
[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a development system for generating an agent from an AI model, running in a computer system, comprising: An AI role setting unit, including a role setting component for labeling the technical field of the AI model and taking the knowledge base of the technical field as the learning knowledge base of the AI model; a goal setting component for setting the learning goal of the AI model and determining the learning endpoint; a work process setting component for determining the logic of the processing event of the AI model; and a constraint condition setting component for defining the boundary of the execution of the AI model; An AI learning direction setting unit for secondary screening of the selected knowledge base, filtering out data not belonging to the learning direction, and making the AI learning database pure; The AI model training unit uses a preset training algorithm to train the initial AI model and generate an AI model with preliminary intelligent agent functions. The functional verification unit constructs verification scenarios and verification indicators corresponding to the functions of the intelligent agent, performs functional verification on the AI model of the preliminary intelligent agent functions, and outputs verification results. The model optimization unit adjusts the constraints of the AI role setting unit based on the verification results and feeds them back to the model training module for secondary training until the model function meets the preset requirements. The model output unit encapsulates AI models that meet preset requirements and outputs deployable AI agents.
[0007] The development principle and advantages of the development system of this invention are as follows: the AI role setting unit completes the identity labeling of the AI model in the specified technical field, avoiding ambiguity in the development direction of the intelligent agent; the AI learning direction setting unit ensures the purity of learning data and reduces training costs; through the closed-loop process of training, verification, and optimization, the intelligent agent achieves efficient iteration and quickly makes up for functional shortcomings; finally, the system outputs an AI intelligent agent in the specified field, which makes the AI intelligent agent highly targeted and adaptable, and can be quickly deployed and applied, significantly reducing the development threshold and cost of functional software.
[0008] Furthermore, in the AI role setting unit, the AI role includes at least one of professional domain role and business scenario role; the completion goal includes quantitative goal and qualitative goal; the workflow includes task decomposition steps, execution order of each step and related logic; the constraints include at least one of data usage scope constraints, output format constraints and performance index constraints.
[0009] Precisely define the application scenarios and functional boundaries of intelligent agents to avoid functional redundancy caused by ambiguous development directions and reduce ineffective development costs; combine quantitative and qualitative goals to ensure that functional targets are quantifiable while also considering flexible needs such as user experience, thereby improving the adaptability of intelligent agents to practical applications.
[0010] Furthermore, the performance metric constraints in the constraint setting component include response time constraints. ; To preset the maximum response time and data processing throughput constraints ; To preset the minimum throughput, a real-time monitoring model is adopted. Dynamic monitoring of performance indicators.
[0011] Furthermore, the data filtering module also integrates the TF-IDF feature extraction algorithm to extract the core features of text data, thereby improving the accuracy of text data filtering.
[0012] Furthermore, the AI learning direction setting unit includes: The direction definition module is used to receive user input of learning direction keywords, domain scope, and data feature descriptions; The data filtering module, based on the information from the direction definition module, employs a feature matching algorithm and a similarity calculation model, using formulas...
[0013] in For data feature vectors, The target direction feature vector is used; the data matching degree is calculated, and data with a matching degree higher than a preset threshold is retained. The database update module stores the filtered valid data into the AI learning database and regularly cleans up expired or invalid data in the database to maintain its purity.
[0014] The secondary screening mechanism significantly reduces the proportion of redundant data in the learning database, reduces data interference during training, and improves model learning efficiency; regular data cleaning ensures the dynamic purity of the database, avoids expired data from affecting the accuracy of model decisions, and reduces the consumption of hardware storage and computing resources.
[0015] Furthermore, the preset training algorithm in the model training unit includes at least one of supervised learning algorithm, reinforcement learning algorithm, and semi-supervised learning algorithm, and a hybrid training model is adopted. ,in For supervised learning models, To reinforce the learning model, It is a semi-supervised learning model. and The model training unit also includes a training parameter configuration module, used to configure training parameters such as learning rate and number of iterations according to the functional requirements of the AI role, wherein the learning rate adopts a dynamically adjusted formula. ; The initial learning rate, The attenuation coefficient is... This represents the number of iterations.
[0016] It integrates multiple pre-set training algorithms such as supervised learning, reinforcement learning, and semi-supervised learning, and adopts a hybrid training model to improve training adaptability. Through the training parameter configuration module, parameters such as learning rate and number of iterations can be dynamically configured according to the functional requirements of the AI role, with the learning rate optimized using a dynamic adjustment formula. Multiple algorithm adaptations meet the training needs of different scenarios, avoiding the limitations of a single algorithm; dynamic parameter configuration enables targeted training, reduces the number of invalid iterations, shortens the model training cycle, and improves the functional accuracy of the model after convergence.
[0017] Furthermore, the verification metrics of the functional verification unit include at least one of the following: functional completion accuracy, response time, data processing efficiency, and stability, employing a multi-dimensional verification model. ; For accuracy metrics, For response time metrics, To process efficiency metrics, As a stability indicator, The verification scenarios include various scenarios such as simulated real-world application scenarios, extreme working conditions scenarios, and boundary condition scenarios.
[0018] Multi-scenario coverage ensures the stability of the agent in complex real-world environments, avoiding functional vulnerabilities caused by single-scenario verification; multi-index verification comprehensively evaluates the agent's performance, providing precise direction for subsequent optimization and reducing the risk of failure after agent deployment; automated verification framework improves verification efficiency and shortens the development cycle.
[0019] Furthermore, the functional verification unit adopts an automated verification framework, which implements parallel testing of multiple scenarios by writing Python scripts. The test case generation adopts an automatic generation model based on genetic algorithms to ensure the coverage, completeness, and diversity of test cases.
[0020] Furthermore, the model optimization unit also includes an optimization suggestion generation module, which generates targeted constraint adjustment suggestions or data filtering rule optimization suggestions based on the functional shortcomings in the verification results, using the gradient descent optimization algorithm, for user reference or automatic adjustment.
[0021] Finally, the encapsulation process of the model output unit includes generating model call interfaces, writing model usage documents, and configuring deployment environment dependency packages. The output deployable AI agent supports local server deployment, cloud server deployment, or edge device deployment, and uses containerized encapsulation technology Docker for environment isolation and dependency management. Attached Figure Description
[0022] Figure 1 System block diagram for building an AI model-generated intelligent agent development system. Detailed Implementation
[0023] The following detailed description illustrates the specific implementation method: Example 1 Appendix Figure 1 As shown, a development system for building AI models to generate intelligent agents is constructed in a computer system. The development system includes an AI role setting unit, an AI learning direction setting unit, an AI model training unit, a functional verification unit, a model optimization unit, and a model output unit. The specific settings of each unit are as follows: The AI role setting unit includes a role setting component for defining roles in professional fields or business scenarios, and using the knowledge base of that technical field as the learning knowledge base for the AI model. A goal setting component sets quantitative and qualitative goals to define the learning objectives of the AI model and determine the learning endpoint. A workflow setting component sets the task breakdown steps, the execution order of each step, and the associated logic to determine the logic for the AI model's processing events. A constraint setting component sets the constraints, including at least one of data usage scope constraints, output format constraints, and performance indicator constraints, wherein the performance indicator constraints utilize a real-time monitoring model. Response time and throughput are dynamically monitored.
[0024] The direction definition subunit within the AI learning direction setting unit receives user input of learning direction keywords, domain scope, and data feature descriptions. In the data filtering subunit, based on the information from the direction definition subunit, a feature matching algorithm and similarity calculation model are used. The data matching degree is calculated using the cosine similarity formula, retaining data with a matching degree higher than a preset threshold. Simultaneously, a TF-IDF feature extraction algorithm is integrated to improve the accuracy of text-based data filtering. The database update subunit stores the filtered valid data into the AI learning database and periodically cleans up expired or invalid data to maintain database purity.
[0025] The AI model training unit incorporates at least one pre-defined training algorithm from supervised learning, reinforcement learning, and semi-supervised learning, employing a hybrid training model. Training is performed; a training parameter configuration module is added to the AI model training unit to configure training parameters such as learning rate and number of iterations according to the functional requirements of the AI role, wherein the learning rate adopts a dynamic adjustment formula. optimization.
[0026] The functional verification unit uses at least one of the following as verification metrics: accuracy, response time, data processing efficiency, and stability, and employs a multi-dimensional verification model. A comprehensive evaluation is conducted; multiple scenarios, including simulated real-world application scenarios, extreme working conditions scenarios, and boundary condition scenarios, are added to the functional verification unit. An automated verification framework is used to achieve parallel testing of multiple scenarios, and test cases are generated through an automatic generation model based on a genetic algorithm.
[0027] The model optimization unit adjusts the constraints of the AI role setting unit based on the verification results and feeds them back to the model training module for secondary training until the model's functionality meets the preset requirements. The optimization suggestion generation subunit within the model optimization unit uses the gradient descent optimization algorithm based on the functional shortcomings identified in the verification results. Generate targeted suggestions for adjusting constraints or optimizing data filtering rules for users to refer to or for automatic adjustments.
[0028] The model output unit encapsulates AI models that meet preset requirements and outputs deployable AI agents. The encapsulation process includes generating model call interfaces, writing model usage documentation, configuring deployment environment dependency packages, and using Docker containerization technology for environment isolation and dependency management. The output deployable AI agents support deployment on local servers, cloud servers, or edge devices.
[0029] Example 2 In this embodiment, an "e-commerce women's clothing product recommendation intelligent agent" is developed using the development system of this application. The specific process is as follows: AI Role Setting Unit Operation: The AI role is set as an "e-commerce product recommendation role," clearly defining its technical identity as a personalized recommendation agent in a women's apparel e-commerce scenario. The objectives are quantitative (recommended product click-through rate ≥ 15%, user conversion rate ≥ 8%) and qualitative (recommendation results match user historical preferences, products match user needs). The workflow is as follows: receive user behavior data (browsing, favorites, purchase records) → analyze user preference characteristics → match product database data → generate a personalized recommendation list → output recommendation results. Constraints are: use only compliant data from within the e-commerce platform, recommendation list length is limited to 10-20 items, response time ≤ 500 milliseconds, and real-time monitoring model is used. Response time is dynamically monitored.
[0030] The AI learning direction setting unit operates as follows: The direction definition subunit receives user input of the learning direction keywords "e-commerce product recommendation", the domain scope "women's clothing category", and data feature descriptions "user behavior data (browsing duration, number of favorites, purchase history) and product data (category, price, style, sales volume)". The data filtering subunit uses a cosine similarity algorithm to filter the raw data, retaining data with a matching degree higher than 85% with the "women's clothing product recommendation" direction, and removing irrelevant category data such as home appliances and digital products, as well as invalid user behavior data. At the same time, the TF-IDF feature extraction algorithm is used to optimize the filtering accuracy of product text description data. The database update subunit stores the filtered valid data into the AI learning database according to the categories of "user data - product data - behavior-related data", and is set to clean up historical data that has not been updated for more than 3 months once a month.
[0031] Model training unit operations: Reinforcement learning was selected as the core training algorithm. A hybrid training model was configured with α=0.2, β=0.7, and γ=0.1. The initial learning rate η0=0.001, iteration count 1000, and batch size 64 were configured via the training parameter configuration module. The learning rate was dynamically adjusted using the formula... Optimization was carried out; based on the AI role setting results and women's clothing-related user behavior data and product data in the clean database, the initial recommendation model was trained in stages. The first stage trained the user preference feature extraction capability, and the second stage trained the product matching accuracy to generate a preliminary product recommendation AI agent.
[0032] Functional verification unit operation: Verification scenarios were constructed, including daily shopping scenarios, major promotional event scenarios, new user scenarios, and niche product scenarios. Verification metrics were set as recommendation click-through rate, user conversion rate, response time, and recommendation diversity. A multi-dimensional verification model was used with ω1=0.4, ω2=0.2, ω3=0.2, and ω4=0.2 for comprehensive evaluation. Parallel automated testing across multiple scenarios was implemented using Python scripts. Test cases were automatically generated using a genetic algorithm. The verification results were: recommendation click-through rate 12%, user conversion rate 6%, response time 400 milliseconds, and recommendation diversity met requirements. However, click-through rate and conversion rate did not meet the preset targets.
[0033] Model optimization unit operations: Analyzing the verification results, the functional weakness was identified as insufficient accuracy in user preference feature analysis, failing to fully consider key dimensions such as user age and consumption level; the optimization suggestion generation subunit uses the gradient descent optimization algorithm to generate optimization suggestions: adjusting constraints, adding "user profile refinement dimensions (age, consumption level, region)", optimizing data filtering rules, and increasing the weight of strongly correlated data such as "purchase after collection" and "multiple browsing" in user behavior data; the above adjustments are automatically executed and fed back to the model training unit for secondary training.
[0034] Model output unit operations: After secondary training, the agent was validated, achieving a recommendation click-through rate of 16% and a user conversion rate of 9%, with all indicators meeting the preset requirements. The agent was encapsulated to generate a RESTful model call interface, and usage documentation including interface parameter descriptions, call examples, and exception handling solutions was written. Environment dependencies such as Python 3.8+ and TensorFlow 2.5+ were configured, and Docker containerization technology was used for environment isolation. A deployable e-commerce women's clothing product recommendation AI agent was output, supporting deployment on the e-commerce platform's cloud server.
[0035] Example 3 In this embodiment, a "medical medical record text classification intelligent agent" is developed using the development system of this application. The specific process is as follows: AI Role Setting Unit Operation: The AI role is set as a "Medical Case Text Classification Role," clearly defining its technical field as an intelligent agent for the automatic classification of cases from common departments in general hospitals. The objectives are quantitative (case classification accuracy ≥98%, classification error rate ≤2%) and qualitative (accurately distinguishing 10 categories of cases, including internal medicine, surgery, and obstetrics and gynecology). The workflow is: receiving case text data → text preprocessing (word segmentation, stop word removal) → extracting text features → performing category classification → outputting classification results. Constraints are: using only compliant medical case data, outputting classification results must include confidence levels, processing time per case ≤300 milliseconds, and real-time monitoring of the model. Dynamically monitor processing time.
[0036] The AI learning direction setting unit operates as follows: The direction definition subunit receives user input of the learning direction keywords "medical case text classification", the domain scope "common departments in general hospitals", and the data feature description "key information such as symptom descriptions, examination results, and diagnostic conclusions in the case text"; The data filtering subunit uses the TF-IDF feature extraction algorithm and the support vector machine classification model to filter the raw data, retaining data with a relevance of more than 90% to medical case classification, and removing non-case text data and invalid case data; The database update subunit stores the valid data into the AI learning database, establishes data labels by department category, and sets up the cleanup of expired case data every quarter.
[0037] Model training unit operations: The Convolutional Neural Network (CNN) algorithm from supervised learning was selected as the core training algorithm. A hybrid training model was configured with α=0.8, β=0.1, and γ=0.1. The initial learning rate η0=0.0005, the number of iterations was set to 800, and the batch size was set to 32 via the training parameter configuration module. The learning rate was dynamically adjusted using the formula... Optimization was carried out; based on the AI role setting results and case data in the clean database, the initial text classification model was trained, focusing on strengthening the association training between symptom features and department categories, and generating a preliminary case text classification AI agent.
[0038] Functional verification unit operation: Verification scenarios were constructed, including routine case scenarios, rare symptom case scenarios, and multi-departmental cross-case scenarios. Verification metrics were set as classification accuracy, classification error rate, processing time, and confidence level achievement rate. A multi-dimensional verification model was used with ω1=0.5, ω2=0.2, ω3=0.1, and ω4=0.2 for comprehensive evaluation. Parallel automated testing of multiple scenarios was implemented using Python scripts. Test cases were automatically generated using a genetic algorithm to verify the initial agent. The verification results were: classification accuracy 96%, classification error rate 3%, processing time 250 milliseconds, and confidence level achievement rate 95%. The classification accuracy did not meet the preset target.
[0039] Model optimization unit operations: Analyze the verification results and determine that the functional shortcoming is insufficient feature extraction of rare symptom cases; The optimization suggestion generation sub-unit uses the gradient descent optimization algorithm to generate optimization suggestions: Adjust the constraints, add the "rare symptom keyword library matching" constraint, optimize the data screening rules, increase the screening weight of rare case data and supplement relevant data; After the user refers to the optimization suggestions, the adjustment operation is executed and fed back to the model training unit for secondary training.
[0040] Model output unit operations: After secondary training, the model classification accuracy reaches 98.5%, and all indicators meet the preset requirements; the model is encapsulated, an RPC call interface is generated, detailed deployment and usage documentation is written, relevant environment dependencies are configured, and Docker containerization technology is used for environment isolation; a deployable medical case text classification AI agent is output, which can be deployed on the hospital's local server to ensure data security.
[0041] The above are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art, knowing all the common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application shall be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A development system for generating intelligent agents from AI models, running on a computer system, characterized in that, include: The AI role setting unit includes a role setting component, which is used to define the technical field of the AI model and use the knowledge base of that technical field as the learning knowledge base of the AI model. The goal setting component is used to set learning goals for the AI model and determine the learning endpoint; the workflow setting component is used to determine the logic for the AI model to process events; and the constraint setting component is used to define the boundaries of the AI model's execution. The AI learning direction setting unit performs a secondary screening of the selected knowledge base, filtering out data that does not belong to the learning direction, thus making the AI learning database pure. The AI model training unit uses a preset training algorithm to train the initial AI model and generate an AI model with preliminary intelligent agent functions. The functional verification unit constructs verification scenarios and verification indicators corresponding to the functions of the intelligent agent, performs functional verification on the AI model of the preliminary intelligent agent functions, and outputs verification results. The model optimization unit adjusts the constraints of the AI role setting unit based on the verification results and feeds them back to the model training module for secondary training until the model function meets the preset requirements. The model output unit encapsulates AI models that meet preset requirements and outputs deployable AI agents.
2. The development system according to claim 1, characterized in that, In the AI role setting unit, the AI role includes at least one of professional domain role and business scenario role; the completion goal includes quantitative goal and qualitative goal; the workflow includes task decomposition steps, execution order of each step and related logic; the constraints include at least one of data usage scope constraints, output format constraints and performance index constraints.
3. The development system according to claim 2, characterized in that, The performance constraints in the constraint setting component include response time constraints. ; To preset the maximum response time and data processing throughput constraints ; To preset the minimum throughput, a real-time monitoring model is adopted. Dynamic monitoring of performance indicators.
4. The development system according to claim 3, characterized in that, The data filtering module also integrates the TF-IDF feature extraction algorithm to extract the core features of text data, thereby improving the accuracy of text data filtering.
5. The development system according to claim 1, characterized in that, The AI learning direction setting unit includes: The direction definition module is used to receive user input of learning direction keywords, domain scope, and data feature descriptions; The data filtering module, based on the information from the direction definition module, employs a feature matching algorithm and a similarity calculation model, using formulas... in For data feature vectors, The target direction feature vector is used; the data matching degree is calculated, and data with a matching degree higher than a preset threshold is retained. The database update module stores the filtered valid data into the AI learning database and regularly cleans up expired or invalid data in the database to maintain its purity.
6. The development system according to claim 1, characterized in that, The preset training algorithm in the model training unit includes at least one of supervised learning algorithm, reinforcement learning algorithm, and semi-supervised learning algorithm, and a hybrid training model is adopted. ,in For supervised learning models, To reinforce the learning model, It is a semi-supervised learning model. and The model training unit also includes a training parameter configuration module, used to configure training parameters such as learning rate and number of iterations according to the functional requirements of the AI role, wherein the learning rate adopts a dynamically adjusted formula. ; The initial learning rate, The attenuation coefficient is... This represents the number of iterations.
7. The development system according to claim 1, characterized in that, The verification metrics of the functional verification unit include at least one of the following: functional completion accuracy, response time, data processing efficiency, and stability, and a multi-dimensional verification model is adopted. ; For accuracy metrics, For response time metrics, To process efficiency metrics, As a stability indicator, The verification scenarios include various scenarios such as simulated real-world application scenarios, extreme working conditions scenarios, and boundary condition scenarios.
8. The development system according to claim 7, characterized in that, The functional verification unit adopts an automated verification framework, which implements parallel testing of multiple scenarios by writing Python scripts. The test case generation adopts an automatic generation model based on genetic algorithms to ensure the coverage, completeness and diversity of test cases.
9. The development system according to claim 1, characterized in that, The model optimization unit also includes an optimization suggestion generation module, which generates targeted constraint adjustment suggestions or data filtering rule optimization suggestions based on the functional shortcomings in the verification results, using the gradient descent optimization algorithm, for user reference or automatic adjustment.
10. The system according to claim 1, characterized in that, The encapsulation process of the model output unit includes generating model call interfaces, writing model usage documents, and configuring deployment environment dependency packages. The output deployable AI agent supports local server deployment, cloud server deployment, or edge device deployment, and uses Docker containerization encapsulation technology for environment isolation and dependency management.