Customizable ai agent management platform for enterprise needs and management method thereof

Abstract

The application discloses an enterprise-demand-oriented customizable AI intelligent agent management platform and a management method thereof, and relates to the technical field of artificial intelligence. The platform comprises a demand acquisition module, a demand analysis module, an application scenario decomposition module, an AI Agent deployment module and a monitoring feedback module. Enterprise demand description data is collected for preprocessing, demand elements are extracted to be converted into execution tasks, and an execution task sequence is constructed. The execution task sequence is decomposed based on an application scenario to construct a sub-task execution chain adapted to the demand application scenario. An intelligent agent recommendation model is constructed, a model is selected from a pre-training model library according to the sub-task execution chain, and an AI Agent of the application scenario is configured and deployed. The running effect is evaluated according to the execution result of the AI Agent, and optimization and improvement are performed. The application flexibly configures the AI Agent according to the enterprise demand, supports multi-scene application, reduces the threshold for enterprises to use AI technology, and improves the application effect of AI technology in actual business.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a customizable AI agent management platform and management method for enterprise needs. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, its application scenarios across various industries are becoming increasingly widespread. From intelligent manufacturing and financial risk control to medical diagnosis and customer service, AI technology is profoundly changing the operating models and service methods of enterprises. However, enterprises often face numerous challenges when introducing AI technology: High technical barriers: Enterprises need professional AI technology teams to develop, train, and deploy AI models, which is costly for many small and medium-sized enterprises; Difficult scenario adaptation: Different enterprises have significantly different business needs and application scenarios, and general AI solutions are difficult to meet personalized needs, resulting in poor implementation effects of AI technology; Complex resource integration: The application of AI technology involves multiple stages such as data collection, model training, computing resource allocation, and system integration, requiring enterprises to invest a lot of time and resources in coordination and management; Insufficient dynamic adjustment: Enterprise business needs and technological environments may change at any time, and traditional AI solutions lack flexibility and dynamic adjustment capabilities, making it difficult to respond quickly to changes.

[0003] Currently, some AI platforms have attempted to address these issues, such as cloud-based AI service platforms and open-source AI frameworks. However, these platforms often focus on technical support and lack a deep understanding of enterprise business needs and the ability to adapt to specific scenarios. Therefore, building an AI service management platform has become an urgent need in current technological development. This platform aims to help enterprises quickly configure and deploy AI Agents (intelligent agents) according to their own needs in a modular and intelligent manner, and flexibly apply them to different business scenarios. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a customizable AI agent management platform and its management method tailored to enterprise needs. This platform enables flexible configuration of AI agents based on enterprise requirements, supports multi-scenario applications, lowers the barrier to entry for enterprises using AI technology, and enhances the effectiveness of AI technology in practical business operations.

[0005] The first aspect of this invention provides a customizable AI agent management platform for enterprise needs, including a requirement acquisition module, a requirement parsing module, an application scenario decomposition module, an AI agent deployment module, and a monitoring and feedback module;

[0006] The requirement acquisition module is responsible for collecting enterprise requirement description data in multimodal form and preprocessing the enterprise requirement description data;

[0007] The requirement parsing module is responsible for extracting requirement elements from the preprocessed enterprise requirement description data, transforming the requirement elements into execution tasks, and constructing an execution task sequence.

[0008] The application scenario decomposition module is responsible for decomposing the execution task sequence based on the application scenario, generating sub-tasks, and constructing a sub-task execution chain adapted to the application scenario.

[0009] The AI ​​Agent deployment module is responsible for building an intelligent agent recommendation model, selecting models from the pre-trained model library according to the sub-task execution chain of different execution tasks in the application scenario, calling the pre-trained models for combination and integration, and configuring and deploying the AI ​​Agent for the application scenario.

[0010] The monitoring and feedback module is responsible for monitoring the running status and performance of the AI ​​Agent in real time, collecting user feedback data, evaluating the running effect of the AI ​​Agent based on the execution results of the AI ​​Agent, and optimizing and improving it based on user feedback data.

[0011] In this solution, the requirement acquisition module collects enterprise requirement description data in multimodal form and preprocesses the enterprise requirement description data, specifically as follows:

[0012] Obtain the target company's demand text description data and demand image description data, and perform data standardization processing on the demand text description data and demand image description data respectively;

[0013] The standardized text and image description data of the requirements are encoded using ALBERT and MobileNetV3 pre-trained models to obtain the corresponding text and image information representations.

[0014] The text information representation and image information representation are subjected to intramodal regularization feature representation using L1 and L2 regularization. The regularized multimodal features are then concatenated to obtain the fused information representation after cross-modal fusion.

[0015] A multimodal attention mechanism is introduced to assign attention weights to text information representation and image information representation, enhance the relevant information representation between different modal data, perform fine-grained semantic alignment, and use global average pooling to filter out data whose attention weights do not meet the preset threshold, so as to obtain the semantically aligned matching information representation.

[0016] The fused information representation after cross-modal fusion and the matching information representation after semantic alignment are aggregated to obtain the requirement description representation set corresponding to the target enterprise's requirements.

[0017] In this solution, the requirement parsing module extracts requirement elements from the preprocessed enterprise requirement description data, transforms these requirement elements into execution tasks, and constructs an execution task sequence, specifically as follows:

[0018] A demand description representation set is generated after preprocessing the enterprise demand description data. A dual-channel neural network model is constructed using CNN and Bi-LSTM networks, and the information representation in the demand description representation set is used as the model input.

[0019] In the CNN network channel, the information representation is convolved using convolution kernels of different sizes to obtain different feature vectors. Max pooling is then performed on the feature vectors, and the feature vectors are concatenated using fully connected layers to obtain local semantic feature vectors.

[0020] In the Bi-LSTM network channel, bidirectional semantic learning is performed on the information representation through a gating mechanism to obtain forward semantic information and backward semantic information. The forward semantic information and backward semantic information are fused to obtain a global semantic feature vector with contextual semantic information representation.

[0021] The local semantic feature vector is concatenated with the global semantic feature vector, and a fully connected layer is used to obtain the probability score of each demand element label. The probability scores of each demand element label are then imported into a conditional random field to learn constraints for labeling and decoding, thereby identifying demand elements.

[0022] Extract the demand elements of the target enterprise to determine the application scenario, construct the execution tasks corresponding to the demand elements according to the application scenario, and obtain the corresponding execution task sequence based on the positional order of the demand elements.

[0023] In this solution, the application scenario decomposition module decomposes the execution task sequence based on the application scenario, generates sub-tasks, and constructs a sub-task execution chain adapted to the application scenario, specifically as follows:

[0024] Based on the application scenario and execution task sequence, and combined with big data methods, task instances that can achieve the required elements are retrieved. The execution data of the task instances are then decomposed into corresponding sub-tasks to generate sub-task sets for different task instances.

[0025] The intersection of the subtask sets of different task instances is performed to obtain the necessary subtasks for different execution tasks. Based on the necessary subtasks, the improved firefly algorithm is used to decompose each execution task.

[0026] In the subtask sets of different task instances, necessary subtasks are randomly combined to generate an initial firefly population. A fitness function is constructed by calculating the task granularity and task cohesion of the subtasks. The brightness of each firefly at different locations is obtained according to the fitness function.

[0027] For each firefly, it moves towards the firefly with higher fitness based on the distance traveled. During the movement, the Levy flight strategy is introduced to update the position. When the new position is reached, the fitness is calculated and the fitness of individual fireflies is compared and judged again.

[0028] When the maximum number of iterations is reached or the convergence condition is met, the optimal combination of subtasks for executing the task is obtained based on the best position of the firefly. The relationship between subtasks in the optimal combination of subtasks is obtained based on the task instance. The subtask execution chain for executing the task is generated through the relationship tags between subtasks.

[0029] In this solution, the AI ​​Agent deployment module constructs an intelligent agent recommendation model, selecting models from a pre-trained model library based on the sub-task execution chains of different tasks in the application scenario. Specifically:

[0030] Automatic entity extraction is performed by obtaining subtask information and pre-trained models used in task instances. A pre-trained model library is built using the extracted pre-trained models. An interaction matrix is ​​constructed based on whether there are interaction records between different subtask execution chains and each pre-trained model. A domain knowledge graph is constructed according to the triple structure using the extracted entities and interaction matrix.

[0031] The sub-task execution task chains corresponding to different enterprise execution tasks are imported into the domain knowledge graph for localization. Random walks are used to obtain sampling paths containing all sub-task nodes in the sub-task execution task chains and recommended pre-trained models. Based on the obtained sampling paths, an execution task-pre-trained model subgraph is constructed.

[0032] The concept of federation is introduced to improve the graph attention network. An agent recommendation model is constructed. The graph attention network is used to obtain the subgraph embedding representation of each execution task-pre-trained model subgraph. The subgraph embedding representation obtained by the local graph attention network is uploaded. The global graph attention network is used to aggregate and update the subgraph embedding representation, and then it is resent to the local graph attention network.

[0033] A multilayer perceptron is added after the graph attention network as the output layer of the agent recommendation model. The updated subgraph embedding representation is obtained and concatenated with the embedding representation of the subtask execution task chain in the knowledge graph. The result is then imported into the multilayer perceptron and the sigmoid function is used to predict the score. The best pre-trained model is recommended for each execution task based on the score.

[0034] In this solution, the AI ​​Agent deployment module calls pre-trained models for combination and integration, and configures and deploys the AI ​​Agent for the application scenario, specifically as follows:

[0035] The system retrieves the best pre-trained model corresponding to each execution task from the pre-trained model library, loads the pre-trained model using API or tools, checks whether the input and output formats of the pre-trained model match the business data, and completes the loading of the pre-trained model when the formats match.

[0036] The pre-trained model is fine-tuned based on the business data of the task, then deployed to the application scenario. The input and output interfaces of the model are configured, the pre-trained model is combined, integrated with the business system, and the AI ​​Agent for the application scenario is deployed.

[0037] In this solution, the monitoring and feedback module evaluates the AI ​​Agent's performance based on its execution results and optimizes and improves it based on user feedback data. Specifically:

[0038] Business keywords are extracted from business data in the application scenario, a retrieval task is constructed based on the business keywords, and business execution data that meets the retrieval requirements is obtained from the task instance using the retrieval task. The business execution data is then subjected to structured preprocessing.

[0039] The sub-task execution chain of the execution task is filled with business execution data, and a business dataset with execution task labels is constructed. In the business dataset, the business execution data with the highest relevance to business needs is obtained according to the Pearson correlation coefficient. The remaining business execution data of the business dataset are introduced one by one. The mRMR algorithm is used to obtain the importance score of the business execution data based on the principle of maximizing the relevance between the business execution data and the performance evaluation and minimizing the redundancy between the business execution data.

[0040] Based on the importance score, a preset number of business execution data are selected to generate an evaluation index for the AI ​​Agent's performance. The evaluation index is used to obtain index parameters from the AI ​​Agent's execution results, and the standard index parameters under ideal conditions are obtained based on the evaluation index.

[0041] The autoencoder network is trained using the standard index parameters of the evaluation indicators. The obtained index parameters are then imported into the trained autoencoder network for data reconstruction to obtain the index parameter prediction values ​​of each evaluation indicator under ideal conditions.

[0042] Calculate the residual value between the estimated value of the indicator parameter and the indicator parameter, use the residual value to determine the anomaly evaluation indicator, determine whether the AI ​​Agent meets the business needs based on the anomaly evaluation indicator and user feedback data, and adjust the AI ​​Agent parameters or select other pre-trained models to reconstruct the AI ​​Agent based on the judgment result.

[0043] The second aspect of this invention discloses a customizable AI agent management method tailored to enterprise needs, comprising the following steps:

[0044] Collect enterprise demand description data in multimodal form, and preprocess the enterprise demand description data;

[0045] Extract demand elements from the preprocessed enterprise demand description data, transform the demand elements into execution tasks, and construct an execution task sequence;

[0046] The execution task sequence is decomposed based on the application scenario to generate sub-tasks, and a sub-task execution chain adapted to the application scenario is constructed.

[0047] Construct an intelligent agent recommendation model, select a model from the pre-trained model library based on the sub-task execution chain of different execution tasks in the application scenario, call the pre-trained model for combination and integration, and configure and deploy the AI ​​Agent for the application scenario;

[0048] Monitor the running status and performance of the AI ​​Agent in real time, collect user feedback data, evaluate the running effect of the AI ​​Agent based on the execution results, and optimize and improve it based on user feedback data.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention constructs a customizable AI agent management platform tailored to enterprise needs, lowering the barrier for enterprises to use AI technology. The platform provides modular AI Agent configuration and pre-trained model library, enabling enterprises to quickly deploy AI solutions without having a professional AI technology team. It provides enterprises with full-process support from needs analysis to AI Agent configuration, deployment, monitoring and optimization.

[0051] Furthermore, the platform supports flexible adaptation to multiple scenarios. By providing various pre-configured AI Agent templates, it allows enterprises to flexibly adjust according to specific business scenarios (such as intelligent customer service, risk warning, production optimization, etc.). This enables the application of AI technology to multiple business processes, maximizing the value of AI. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0053] Figure 1A block diagram of a customizable AI agent management platform tailored to enterprise needs is shown.

[0054] Figure 2 The flowchart illustrating the extraction of requirement elements and construction of an execution task sequence in this embodiment is shown.

[0055] Figure 3 The flowchart illustrates the process of selecting a model from a pre-trained model library to build an agent recommendation model;

[0056] Figure 4 A flowchart illustrating a customizable AI agent management approach tailored to enterprise needs is provided. Detailed Implementation

[0057] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0059] Figure 1 A block diagram of a customizable AI agent management platform tailored to enterprise needs is shown.

[0060] like Figure 1 As shown, the first embodiment of the present invention provides a customizable AI agent management platform for enterprise needs, including: a requirement acquisition module 101, a requirement parsing module 102, an application scenario decomposition module 103, an AI Agent deployment module 104, and a monitoring and feedback module 105.

[0061] The requirement acquisition module 101 is responsible for collecting enterprise requirement description data in multimodal form and preprocessing the enterprise requirement description data;

[0062] The requirement parsing module 102 is responsible for extracting requirement elements from the preprocessed enterprise requirement description data, converting the requirement elements into execution tasks, and constructing an execution task sequence.

[0063] The application scenario decomposition module 103 is responsible for decomposing the execution task sequence based on the application scenario, generating sub-tasks, and constructing a sub-task execution chain adapted to the application scenario.

[0064] The AI ​​Agent deployment module 104 is responsible for building an intelligent agent recommendation model, selecting a model from the pre-trained model library according to the sub-task execution chain of different execution tasks in the application scenario, calling the pre-trained model for combination and integration, and configuring and deploying the AI ​​Agent for the application scenario.

[0065] The monitoring and feedback module 105 is responsible for monitoring the running status and performance of the AI ​​Agent in real time, collecting user feedback data, evaluating the running effect of the AI ​​Agent based on the execution results of the AI ​​Agent, and optimizing and improving it based on user feedback data.

[0066] It should be noted that the demand acquisition module collects and analyzes the business needs of enterprises, transforming them into specific AI technology requirements. Through methods such as questionnaires, interactive dialogues, or data import, textual and image descriptions of the target enterprise's needs are obtained. This clarifies the pain points, target scenarios, and expected results (such as classification, prediction, and generation) that the enterprise needs to address. The textual and image descriptions of the demands are then subjected to data standardization processes such as cleaning, labeling, and format conversion to ensure that the data quality meets the requirements of the AI ​​model.

[0067] The standardized text and image description data of the requirements are encoded using ALBERT and MobileNetV3 pre-trained models. A linear layer is added to the MobileNetV3 pre-trained model to make the dimension of the image hidden layer consistent with the dimension of the text features, thereby obtaining the corresponding text and image information representations. The text and image information representations are then subjected to intra-modal regularization using L1 and L2 regularization. L1 regularization is used in the linear layer to improve sparsity and achieve feature selection by adding a penalty term proportional to the sum of the absolute values ​​of the weights. L2 regularization reduces complexity and prevents overfitting by adding a penalty term proportional to the sum of the squares of the weights. The regularized multimodal features are concatenated to obtain the fused information representation after cross-modal fusion. A multimodal attention mechanism is introduced to assign attention weights to textual and image information representations, enhancing the representation of relevant information between different modalities and performing fine-grained semantic alignment. The aligned and matched features can represent the relevant information between the two modalities. The larger the attention weight, the higher the attention the corresponding information representation receives. Global average pooling is used to filter out data whose attention weight does not meet the preset threshold to obtain the semantically aligned matching information representation. The fused information representation after cross-modal fusion and the semantically aligned matching information representation are aggregated to obtain the requirement description representation set corresponding to the target enterprise's requirements.

[0068] Figure 2 A flowchart illustrating the extraction of requirement elements and the construction of an execution task sequence is shown in this embodiment.

[0069] According to an embodiment of the present invention, in the requirement parsing module, requirement elements are extracted from the preprocessed enterprise requirement description data, the requirement elements are transformed into execution tasks, and an execution task sequence is constructed, specifically as follows:

[0070] S202, Obtain the demand description representation set generated after preprocessing the enterprise demand description data, construct a dual-channel neural network model using CNN network and Bi-LSTM network, and use the information representation in the demand description representation set as the model input;

[0071] S204, In the CNN network channel, the information representation is convolved using convolution kernels of different sizes to obtain different feature vectors, the feature vectors are subjected to max pooling operation, and the feature vectors are concatenated using fully connected layers to obtain local semantic feature vectors;

[0072] S206, In the Bi-LSTM network channel, bidirectional semantic learning is performed on the information representation through a gating mechanism to obtain forward semantic information and backward semantic information. The forward semantic information and backward semantic information are fused to obtain a global semantic feature vector with contextual semantic information representation.

[0073] S208, the local semantic feature vector is concatenated with the global semantic feature vector, a fully connected layer is used to obtain the probability score of each demand element label, the probability scores of each demand element label are imported into the conditional random field learning constraints for annotation and decoding, and demand element identification is performed.

[0074] S210: Extract the demand elements of the target enterprise to determine the application scenario, construct the execution tasks corresponding to the demand elements according to the application scenario, and obtain the corresponding execution task sequence based on the positional order of the demand elements.

[0075] It should be noted that a dual-channel neural network model using CNN and Bi-LSTM networks is constructed to extract the enterprise's demand elements. Business demands are the core basis for configuring the AI ​​Agent, directly determining its functions and objectives. Demand elements can include specific problems (such as improving efficiency, reducing costs, and optimizing decision-making), required functions (such as classification, prediction, generation, and recommendation), and specific application scenarios (such as intelligent customer service, risk warning, production optimization, and marketing). For example, in e-commerce, the business objective is to improve the accuracy of product recommendations; in manufacturing, the business objective is to optimize production processes. The CNN network channels capture local semantic information of the demand description representation set, while the Bi-LSTM network channels capture global semantic information. The dual-channel structure simultaneously captures features of different granularities, improving the efficiency of demand element identification and extraction. Different sized convolutional kernels are used in the CNN network channels to extract local features of the demand description representation set at multiple levels. Max pooling is then performed on the feature vectors to avoid overfitting caused by excessive subsequent parameters. Because the max pooling operation in CNN networks performs feature selection, the semantic feature extraction of the demand description representation set is not complete enough. Bi-LSTM network channels overcome this problem, avoiding excessive neglect of semantic features. Bi-LSTM is an improved recurrent neural network that can simultaneously capture the forward and backward dependencies of sequence data, thus extracting global semantic features more comprehensively. In the Bi-LSTM network channels, the information representation sequences are processed from front to back and from front to back respectively, generating forward hidden state sequences and backward hidden state sequences, which are then concatenated to obtain the global semantic feature vector.

[0076] Conditional random fields (CRFs) are used to introduce constraints that ensure the validity of the final prediction. These constraints are automatically learned by the CRF layers during training, improving the accuracy of demand element identification and extraction. Furthermore, adversarial training is introduced during the training of the dual-channel neural network model. Perturbation factors are added to generate adversarial examples corresponding to different demand element labels. These adversarial examples, along with the training samples, are input into the model for training. The deviation between the predicted and actual results is calculated, and backpropagation is performed to update the model parameters.

[0077] It should be noted that in the application scenario decomposition module, task instances capable of fulfilling the required elements are retrieved based on the application scenario and execution task sequence, combined with big data methods. The execution data of these task instances is then used for fine-grained decomposition to obtain corresponding subtasks, generating subtask sets for different task instances. The intersection of these subtask sets is then performed to obtain the necessary subtasks for each execution task. Based on these necessary subtasks, the improved firefly algorithm is used to decompose each execution task. Fireflies attract each other based on brightness; lower-brightness fireflies move towards higher-brightness fireflies. Brightness is related to the objective function value, and the moving distance of the fireflies is affected by the attraction force and the random step size. Based on the necessary subtasks in the subtask sets of different task instances, an initial firefly population is generated by randomly combining them. A fitness function is constructed by calculating the task granularity and task cohesion of the subtasks. The brightness of each firefly at different locations is obtained based on this fitness function. The task granularity in the fitness function describes the size, complexity, and execution time of each subtask. The granularity of the subtasks needs to be determined based on the complexity of the task and the processing power of the hardware. Task cohesion refers to the degree of correlation and connection between subtasks. Strong subtask cohesion can improve task execution efficiency and resource utilization, and avoid task duplication and resource waste. The fitness function is expressed as: The complexity of subtask t is represented by the computational and storage requirements of the subtask. This represents the processing power of hardware device g, calculated using the device's computing and storage capabilities. This indicates the relevance between subtask i and subtask j. Indicates the number of subtasks.

[0078] For each firefly, it moves towards a firefly with higher fitness based on the distance traveled. During the movement, a Levy flight strategy is introduced to update its position. Upon reaching the new position, fitness is calculated, and the fitness comparison between individual fireflies is re-evaluated. The firefly movement formula is:

[0079] ;

[0080] in Indicates the location of the fireflies. Indicates the location of other fireflies. Indicates initial attraction, Indicates the light absorption coefficient. This represents the distance between fireflies i and j. Indicates the step size factor. This represents the random step size generated by the Levi flight.

[0081] The long step size characteristic of Lévy flight helps the firefly escape local optima. Short step sizes are used for local search, while long step sizes are used for global exploration. Introducing the Lévy flight strategy can find the global optimum faster. When the maximum number of iterations is reached or the convergence condition is met, the optimal combination of subtasks to be executed based on the firefly's best position is obtained. The relationships between subtasks in the optimal combination are obtained based on the task instance, and the subtask execution chain to be executed is generated through the relationship tags between subtasks.

[0082] Figure 3 The flowchart illustrates the process of selecting a model from a pre-trained model library to build an agent recommendation model;

[0083] According to the claims of this application, in the AI ​​Agent deployment module, an intelligent agent recommendation model is constructed, and a model is selected from a pre-trained model library based on the sub-task execution chain of different execution tasks in the application scenario, specifically as follows:

[0084] S302, Obtain subtask information and pre-trained models used in task instances to automatically extract entities, build a pre-trained model library using the extracted pre-trained models, construct an interaction matrix based on whether there are interaction records between different subtask execution chains and each pre-trained model, and construct a domain knowledge graph according to the triplet structure using the extracted entities and interaction matrix.

[0085] S304, import the sub-task execution task chains corresponding to different execution tasks of the enterprise into the domain knowledge graph for localization, use random walk to obtain sampling paths containing all sub-task nodes in the sub-task execution task chains and recommended pre-trained models, and construct an execution task-pre-trained model subgraph based on the obtained sampling paths;

[0086] S306 introduces the federated concept to improve the graph attention network, constructs an agent recommendation model, uses the graph attention network to obtain the subgraph embedding representation of each execution task-pre-trained model subgraph, uploads the subgraph embedding representation obtained by the local graph attention network, uses the global graph attention network to aggregate and update the subgraph embedding representation, and resends it to the local graph attention network.

[0087] S308 adds a multilayer perceptron as the output layer of the agent recommendation model after the graph attention network. It obtains the updated subgraph embedding representation and concatenates it with the embedding representation of the subtask execution task chain in the knowledge graph. It then imports the sublayer perceptron and uses the sigmoid function to predict the score. Based on the score, it recommends the best pre-trained model for each execution task.

[0088] It should be noted that, based on business requirements and the pre-trained networks used in the task instances, a suitable pre-trained model library (such as TensorFlow Hub, Hugging Face, PyTorch Hub, etc.) should be selected or constructed, and it should be determined whether the pre-trained model library contains suitable pre-trained models (such as BERT, GPT, DeepSeek, llama3, ResNet, etc.). A Node2vec random walk is used to obtain sampling paths in the knowledge graph that include all sub-task nodes in the sub-task execution task chain and the recommended pre-trained models, constructing an execution task-pre-trained model subgraph. A graph attention network is used to obtain the subgraph embedding representations of each subgraph through information passing and neighbor aggregation mechanisms. During the iteration process, the subgraph embedding representations of each local graph attention network are transmitted to the global graph attention network for aggregation; the global graph attention network transmits the updated subgraph embedding representations to each local graph attention network. Through multi-party federated evolution, the feature representation updates of the execution task and candidate pre-trained models are achieved, improving the stability and robustness of the pre-trained recommendations.

[0089] It should be noted that this module retrieves the optimal pre-trained model corresponding to each execution task from the pre-trained model library. It then deploys the trained AI Agent to the specific business scenario of the enterprise, supporting multiple deployment methods (such as cloud, edge computing, local servers, etc.) and dynamically adjusting according to scenario requirements. The module loads the pre-trained model using APIs or tools, checks if the input / output format of the pre-trained model matches the business data, and completes the loading of the pre-trained model when the format matches. It then fine-tunes the pre-trained model based on the business data of the execution task, adjusting parameters such as learning rate, batch size, and number of iterations. The fine-tuned pre-trained model is then deployed to the application scenario, configuring the model's input / output interfaces, combining the pre-trained models, integrating them with the business system, and deploying the AI ​​Agent for the application scenario. The deployed AI Agent is tested to ensure its normal operation. Finally, based on new data and business needs, the model is retrained periodically, and the pre-trained models in the model library are updated to ensure they are always up-to-date.

[0090] It should be noted that in the monitoring and feedback module, business keywords are extracted from business data in the application scenario. A retrieval task is constructed based on these keywords. The retrieval task retrieves business execution data that meets the retrieval requirements from task instances, and performs structured preprocessing on this data. The sub-task execution chain of the execution task is populated based on the business execution data, constructing a business dataset labeled with the execution task. In this business dataset, the business execution data with the highest relevance to business needs is obtained based on the Pearson correlation coefficient. The remaining business execution data from the business dataset are then introduced one by one. The mRMR algorithm is used to obtain an importance score for the business execution data based on the principle of maximizing the relevance between the business execution data and the performance evaluation and minimizing redundancy between the business execution data. Based on the importance score, a preset number of business execution data are selected to generate evaluation metrics for the AI ​​Agent's performance, such as performance metrics (accuracy, speed, response time, throughput, resource consumption, etc.), business value (key performance indicators (indicators directly related to enterprise business objectives), business efficiency improvement, etc.), and technical stability assessment (system availability, fault tolerance, and resource utilization, etc.). The evaluation metrics are used to obtain indicator parameters from the execution results of the AI ​​Agent, and standard indicator parameters under ideal conditions are obtained based on the evaluation metrics. An autoencoder network is trained using the standard indicator parameters of the evaluation metrics. The obtained indicator parameters are then imported into the trained autoencoder network for data reconstruction to obtain the predicted indicator parameter values ​​for each evaluation metric under ideal conditions. The residual values ​​between the predicted indicator parameter values ​​and the actual indicator parameters are calculated. Abnormal evaluation metrics are identified using these residual values. Based on the abnormal evaluation metrics and user feedback data, it is determined whether the AI ​​Agent meets business requirements. Based on the determination results, the AI ​​Agent parameters are adjusted, or other pre-trained models are selected to reconstruct the AI ​​Agent. In this module, the running status and performance of the AI ​​Agent are monitored in real time, user feedback data is collected for model optimization, and the model is periodically updated and retrained based on the evaluation results and feedback results.

[0091] In one embodiment of this application, an enterprise needs an AI Agent capable of automatically answering customer questions; selects a pre-trained BERT model from the Hugging Face model library; selects a suitable BERT variant (such as BERT-base-uncased) for the question-answering task; loads the BERT model using the Hugging Face API; collects customer question data and cleans and labels it; and fine-tunes the BERT model using the customer question data.

[0092] Evaluate the model's accuracy and response time on the test set; quantify the model to improve inference speed; deploy the model to the enterprise's customer service system; monitor the quality of the AI ​​Agent's responses in real time and optimize the model based on user feedback; and regularly update the model to adapt to new customer questions and business needs.

[0093] Figure 4 A flowchart illustrating a customizable AI agent management approach tailored to enterprise needs is shown.

[0094] The second embodiment of the present invention provides a customizable AI agent management method for enterprise needs, applied to a customizable AI agent management platform for enterprise needs, including the following steps:

[0095] S402, Collect enterprise demand description data in multimodal form, and preprocess the enterprise demand description data;

[0096] S404, Extract demand elements from the preprocessed enterprise demand description data, transform the demand elements into execution tasks, and construct an execution task sequence;

[0097] S406, decompose the execution task sequence based on the application scenario, generate sub-tasks, and build a sub-task execution chain adapted to the application scenario.

[0098] S408: Construct an intelligent agent recommendation model, select a model from the pre-trained model library based on the sub-task execution chain of different execution tasks in the application scenario, call the pre-trained model for combination and integration, and configure and deploy the AI ​​Agent for the application scenario.

[0099] The S410 monitors the running status and performance of the AI ​​Agent in real time, collects user feedback data, evaluates the performance of the AI ​​Agent based on its execution results, and optimizes and improves it based on user feedback data.

[0100] The management platform also includes visualization and interactive features, providing enterprise users with a user-friendly visual interface that allows them to configure AI Agents, view results, and analyze reports through graphical operations. The user-friendly interface and interactive design reduce the difficulty of operation while providing real-time feedback and dynamic adjustment capabilities.

[0101] According to an embodiment of the present invention, in the pre-trained model library, multi-dimensional attributes of the pre-trained models are obtained by extracting demand elements, enterprise types, and application scenarios based on historical AI Agent configuration and deployment information, and generating labels for each pre-trained model in the historical AI Agent configuration and deployment process; features corresponding to different attribute labels are extracted by analyzing a large amount of historical AI Agent configuration and deployment information, and an attention mechanism is used to assign weights to the features corresponding to different attribute labels to characterize the importance of different attribute labels; basic information, demand information, and business scenarios of the target enterprise are obtained to comprehensively characterize the target enterprise and construct a target enterprise profile; the target enterprise profile is used to perform matching in the pre-trained model library, and a matching score is generated based on the similarity of attribute labels and the corresponding attribute label weights; pre-trained models that meet the standards are selected based on the matching scores to achieve initial screening; and the best pre-trained model is recommended for each execution task based on the pre-screened pre-trained models, further improving the applicability of the pre-trained models.

[0102] A third embodiment of the present invention provides a computer-readable storage medium including a customizable AI agent management method program for enterprise needs. When the customizable AI agent management method program for enterprise needs is executed by a processor, it implements the steps of a customizable AI agent management method for enterprise needs.

[0103] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, and can be electrical, mechanical, or other forms.

[0104] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] Alternatively, if the integrated modules of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0106] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A customizable AI intelligent agent management platform tailored to enterprise needs, characterized in that: include: The module includes a requirements gathering module, a requirements parsing module, an application scenario decomposition module, an AI Agent deployment module, and a monitoring and feedback module. The requirement acquisition module is responsible for collecting enterprise requirement description data in multimodal form and preprocessing the enterprise requirement description data; The requirement parsing module is responsible for extracting requirement elements from the preprocessed enterprise requirement description data, transforming the requirement elements into execution tasks, and constructing an execution task sequence. The application scenario decomposition module is responsible for decomposing the execution task sequence based on the application scenario, generating sub-tasks, and constructing a sub-task execution chain adapted to the application scenario. The AI ​​Agent deployment module is responsible for building an intelligent agent recommendation model, selecting models from the pre-trained model library according to the sub-task execution chain of different execution tasks in the application scenario, calling the pre-trained models for combination and integration, and configuring and deploying the AI ​​Agent for the application scenario. The monitoring and feedback module is responsible for monitoring the running status and performance of the AI ​​Agent in real time, collecting user feedback data, evaluating the running effect of the AI ​​Agent based on the execution results of the AI ​​Agent, and optimizing and improving it based on user feedback data. In the application scenario decomposition module, the execution task sequence is decomposed based on the application scenario to generate sub-tasks, and a sub-task execution chain adapted to the application scenario is constructed, specifically as follows: Based on the application scenario and execution task sequence, and combined with big data methods, task instances that can achieve the required elements are retrieved. The execution data of the task instances are then decomposed into corresponding sub-tasks to generate sub-task sets for different task instances. The intersection of the subtask sets of different task instances is performed to obtain the necessary subtasks for different execution tasks. Based on the necessary subtasks, the improved firefly algorithm is used to decompose each execution task. In the subtask sets of different task instances, necessary subtasks are randomly combined to generate an initial firefly population. A fitness function is constructed by calculating the task granularity and task cohesion of the subtasks. The brightness of each firefly at different locations is obtained according to the fitness function. For each firefly, it moves towards the firefly with higher fitness based on the distance traveled. During the movement, the Levy flight strategy is introduced to update the position. When the new position is reached, the fitness is calculated and the fitness of individual fireflies is compared and judged again. When the maximum number of iterations is reached or the convergence condition is met, the optimal combination of subtasks for executing the task is obtained based on the best position of the firefly. The relationship between subtasks in the optimal combination of subtasks is obtained based on the task instance. The subtask execution chain for executing the task is generated through the relationship tags between subtasks. In the AI ​​Agent deployment module, an intelligent agent recommendation model is built. Based on the sub-task execution chains of different tasks in the application scenario, a model is selected from the pre-trained model library. Specifically: Automatic entity extraction is performed by obtaining subtask information and pre-trained models used in task instances. A pre-trained model library is built using the extracted pre-trained models. An interaction matrix is ​​constructed based on whether there are interaction records between different subtask execution chains and each pre-trained model. A domain knowledge graph is constructed according to the triple structure using the extracted entities and interaction matrix. The sub-task execution task chains corresponding to different enterprise execution tasks are imported into the domain knowledge graph for localization. Random walks are used to obtain sampling paths containing all sub-task nodes in the sub-task execution task chains and recommended pre-trained models. Based on the obtained sampling paths, an execution task-pre-trained model subgraph is constructed. The concept of federation is introduced to improve the graph attention network. An agent recommendation model is constructed. The graph attention network is used to obtain the subgraph embedding representation of each execution task-pre-trained model subgraph. The subgraph embedding representation obtained by the local graph attention network is uploaded. The global graph attention network is used to aggregate and update the subgraph embedding representation, and then it is resent to the local graph attention network. A multilayer perceptron is added after the graph attention network as the output layer of the agent recommendation model. The updated subgraph embedding representation is obtained and concatenated with the embedding representation of the subtask execution task chain in the knowledge graph. The result is then imported into the multilayer perceptron and the sigmoid function is used to predict the score. The best pre-trained model is recommended for each execution task based on the score.

2. The customizable AI intelligent agent management platform for enterprise needs according to claim 1, characterized in that, In the demand acquisition module, enterprise demand description data in multimodal form is collected, and the enterprise demand description data is preprocessed, specifically as follows: Obtain the target company's demand text description data and demand image description data, and perform data standardization processing on the demand text description data and demand image description data respectively; The standardized text and image description data of the requirements are encoded using ALBERT and MobileNetV3 pre-trained models to obtain the corresponding text and image information representations. The text information representation and image information representation are subjected to intramodal regularization feature representation using L1 and L2 regularization. The regularized multimodal features are then concatenated to obtain the fused information representation after cross-modal fusion. A multimodal attention mechanism is introduced to assign attention weights to text information representation and image information representation, enhance the relevant information representation between different modal data, perform fine-grained semantic alignment, and use global average pooling to filter out data whose attention weights do not meet the preset threshold, so as to obtain the semantically aligned matching information representation. The fused information representation after cross-modal fusion and the matching information representation after semantic alignment are aggregated to obtain the requirement description representation set corresponding to the target enterprise's requirements.

3. The customizable AI intelligent agent management platform for enterprise needs according to claim 1, characterized in that, In the requirements parsing module, requirement elements are extracted from the preprocessed enterprise requirement description data, these requirement elements are transformed into execution tasks, and an execution task sequence is constructed, specifically as follows: A demand description representation set is generated after preprocessing the enterprise demand description data. A dual-channel neural network model is constructed using CNN and Bi-LSTM networks, and the information representation in the demand description representation set is used as the model input. In the CNN network channel, the information representation is convolved using convolution kernels of different sizes to obtain different feature vectors. Max pooling is then performed on the feature vectors, and the feature vectors are concatenated using fully connected layers to obtain local semantic feature vectors. In the Bi-LSTM network channel, bidirectional semantic learning is performed on the information representation through a gating mechanism to obtain forward semantic information and backward semantic information. The forward semantic information and backward semantic information are fused to obtain a global semantic feature vector with contextual semantic information representation. The local semantic feature vector is concatenated with the global semantic feature vector, and a fully connected layer is used to obtain the probability score of each demand element label. The probability scores of each demand element label are then imported into a conditional random field to learn constraints for labeling and decoding, thereby identifying demand elements. Extract the demand elements of the target enterprise to determine the application scenario, construct the execution tasks corresponding to the demand elements according to the application scenario, and obtain the corresponding execution task sequence based on the positional order of the demand elements.

4. The customizable AI intelligent agent management platform for enterprise needs according to claim 1, characterized in that, In the AI ​​Agent deployment module, pre-trained models are called for combination and integration, and AI Agents for deployment scenarios are configured and deployed, specifically as follows: The system retrieves the best pre-trained model corresponding to each execution task from the pre-trained model library, loads the pre-trained model using API or tools, checks whether the input and output formats of the pre-trained model match the business data, and completes the loading of the pre-trained model when the formats match. The pre-trained model is fine-tuned based on the business data of the task, then deployed to the application scenario. The input and output interfaces of the model are configured, the pre-trained model is combined, integrated with the business system, and the AI ​​Agent for the application scenario is deployed.

5. A customizable AI intelligent agent management platform for enterprise needs as described in claim 1, characterized in that, In the monitoring and feedback module, the performance of the AI ​​Agent is evaluated based on its execution results, and optimizations and improvements are made based on user feedback data. Specifically: Business keywords are extracted from business data in the application scenario, a retrieval task is constructed based on the business keywords, and business execution data that meets the retrieval requirements is obtained from the task instance using the retrieval task. The business execution data is then subjected to structured preprocessing. The sub-task execution chain of the execution task is filled with business execution data, and a business dataset with execution task labels is constructed. In the business dataset, the business execution data with the highest relevance to business needs is obtained according to the Pearson correlation coefficient. The remaining business execution data of the business dataset are introduced one by one. The mRMR algorithm is used to obtain the importance score of the business execution data based on the principle of maximizing the relevance between the business execution data and the performance evaluation and minimizing the redundancy between the business execution data. Based on the importance score, a preset number of business execution data are selected to generate an evaluation index for the AI ​​Agent's performance. The evaluation index is used to obtain index parameters from the AI ​​Agent's execution results, and the standard index parameters under ideal conditions are obtained based on the evaluation index. The autoencoder network is trained using the standard index parameters of the evaluation indicators. The obtained index parameters are then imported into the trained autoencoder network for data reconstruction to obtain the index parameter prediction values ​​of each evaluation indicator under ideal conditions. Calculate the residual value between the predicted value of the indicator parameter and the indicator parameter, use the residual value to determine the anomaly evaluation indicator, determine whether the AI ​​Agent meets the business requirements based on the anomaly evaluation indicator and user feedback data, and adjust the AI ​​Agent parameters or select other pre-trained models to reconstruct the AI ​​Agent based on the judgment result.

6. A customizable AI agent management method tailored to enterprise needs, characterized in that: The customizable AI agent management platform for enterprise needs as described in any one of claims 1-5 includes the following steps: Collect enterprise demand description data in multimodal form, and preprocess the enterprise demand description data; Extract demand elements from the preprocessed enterprise demand description data, transform the demand elements into execution tasks, and construct an execution task sequence; The execution task sequence is decomposed based on the application scenario to generate sub-tasks, and a sub-task execution chain adapted to the application scenario is constructed. Construct an intelligent agent recommendation model, select a model from the pre-trained model library based on the sub-task execution chain of different execution tasks in the application scenario, call the pre-trained model for combination and integration, and configure and deploy the AI ​​Agent for the application scenario; Monitor the running status and performance of the AI ​​Agent in real time, collect user feedback data, evaluate the running effect of the AI ​​Agent based on the execution results, and optimize and improve it based on user feedback data.

Images