Application method based on privatized large model and AI intelligent agent

Abstract

The invention provides an application method based on a privatized large model and an AI agent, and the method comprises the steps: deploying a pre-trained large language model on an internal server of an enterprise, carrying out the adjustment of the large language model through the own data of the enterprise, and carrying out the local deployment, and obtaining a privatized large language model; constructing an AI agent framework, and defining a task execution logic, a decision rule and an interactive interface with the privatized large language model of the AI agent; the AI agent receives a task request proposed by a user through a natural language, calls the privatized large language model to analyze the task request, and generates a task execution plan; and the AI agent executes a task according to a result output by the privatized large language model, and feeds back a task execution result to the user. According to the method, a privatized large language model is deployed in an enterprise, and the autonomous decision-making capability of an AI agent is combined, so that the data security is guaranteed, and meanwhile, efficient processing and intelligent decision-making support for complex tasks are realized.

Description

Technical Field

[0001] The present invention relates to the field of AI agent technology, and in particular to an application method based on a privatized large model and an AI agent. Background Art

[0002] With the rapid development of artificial intelligence (AI), large language models (LLMs) have made significant progress in the field of natural language processing (NLP). While powerful, existing large language models (such as the GPT series) are primarily based on public cloud services and have the following limitations: 1) Data security and privacy issues: Enterprise sensitive data needs to be uploaded to the cloud for processing, which poses a risk of leakage.

[0003] 2) Limited customization capabilities: Public models are difficult to deeply customize to meet the needs of specific enterprises.

[0004] 3) High cost: Frequent calls to public models will incur high costs, especially for large-scale enterprise applications.

[0005] Furthermore, AI agents, as an emerging technology, can complete complex workflows through autonomous decision-making and task execution. However, current AI agents mostly rely on public models and lack deep integration with private models, limiting their potential in enterprise applications. Summary of the Invention

[0006] The purpose of this invention is to provide an application method based on the combination of a private large model and an AI agent. By deploying a private large language model within the enterprise and combining it with the autonomous decision-making ability of the AI agent, it can achieve efficient processing of complex tasks and intelligent decision-making support while ensuring data security.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: An application method based on a private large model and an AI agent, the method comprising: Deploy a pre-trained large language model on the company's internal server, adjust the large language model using the company's own data, and deploy it locally to obtain a private large language model. Build an AI agent framework, define the AI agent's task execution logic, decision-making rules, and interaction interface with the private large language model; The AI agent receives task requests from users in natural language, uses a private large language model to analyze the task requests, and generates a task execution plan; The AI agent performs tasks based on the output of the private large language model and feeds back the task execution results to the user.

[0008] Furthermore, before deploying the pre-trained large language model on the enterprise's internal server, the following operations are performed: Obtaining enterprise demand information, including business demand information, technical resource information, and data compliance requirement information; Input enterprise demand information into the preset machine learning model for processing to obtain recommended model architecture information; Determine the large language model architecture for subsequent deployment based on the recommended model architecture information.

[0009] Furthermore, the AI agent receives task requests from users in natural language, uses the large language model to analyze the task requests, and generates a task execution plan, which specifically includes: Receive task requests submitted by users through a natural language interface; Call the private large language model to parse the task request, identify the key requirements and goals of the task request, and generate the task to be executed; Decompose the task to be executed into multiple subtasks and generate a task execution plan; Allocate computing resources according to the task execution plan to complete the execution of subtasks.

[0010] Furthermore, the task to be executed is decomposed into multiple subtasks and a task execution plan is generated, which specifically includes: Analyze the available computing resource information of all computer devices in the enterprise internal network and obtain the available computing resource evaluation results; Decomposing the task to be executed according to the evaluation result of available computing resources to obtain a subtask set including multiple subtasks of unequal sizes; Allocate each subtask in the subtask set to different computer devices in the enterprise internal network according to its own scale to obtain the allocation result; Generate a task execution plan based on the allocation results.

[0011] Furthermore, the available computing resource information of all computer devices in the enterprise internal network is analyzed to obtain the available computing resource evaluation results, including: Determine the number of computer devices in the enterprise's internal network, as well as a sampling method and computing resource indicators, wherein the sampling method is used to limit the collection time period and collection frequency of computing resource indicators; Collect computing resource indicators of all computer devices according to a sampling method, and generate a computing resource matrix reflecting the real-time computing resource usage of the computer devices; The computing resource matrix is input into the neural network model for processing to obtain the prediction results of the available computing resource information of all computer devices in the enterprise internal network in the future period of time as the available computing resource evaluation results.

[0012] Furthermore, during the execution of the subtasks, the execution order of each subtask is dynamically adjusted. Specifically, the dynamic adjustment includes regularly performing the following operations: Determine the execution ranking indicators of the subtasks and normalize the execution ranking indicators; Define the fuzzy set of execution ranking indicators and further define the membership function of the fuzzy set; Set the subtask execution sorting rules, apply the subtask execution sorting rules to the fuzzy set, and output the sorted fuzzy set; Defuzzify the sorting fuzzy set to obtain the specific execution sorting information of each subtask; Adjust the execution order of the subtasks according to the execution sorting information of each subtask.

[0013] Furthermore, when the AI agent analyzes the task request, it calls the private large language model to determine whether the task request contains content that may lead to the leakage of corporate privacy information and holds the relevant users accountable. Specifically, the following operations are performed: The AI agent identifies the identity of the user who initiated the task request, obtains the user identity information, which includes at least the name and position, and determines the user data permissions based on the user identity information; Call the privatized large language model to analyze whether there is any possible confidential content in the task request, determine the data type of the possible confidential content, and send the data type of the confidential data to the privacy proof module. The privacy proof module determines the privacy data features used for proof based on the data type of the possible confidential data; The AI agent and privacy proof module use a zero-knowledge proof algorithm to verify whether the potentially confidential content is private data based on the characteristics of the potentially confidential content and private data. If the potentially confidential content is not private data, the task execution plan is generated. If the potentially confidential content is private data, the privacy proof module determines the data permissions required to use the private data, and combines the user's data permissions to determine whether the user has violated the data usage regulations. If so, an accountability alarm message is generated.

[0014] Furthermore, the privacy proof module determines the data permissions required to use the private data and, based on the user's data permissions, determines whether the user has engaged in illegal data usage. Specifically, the module: Determine whether the user's data permissions meet the data permissions required to use private data. If so, further determine whether the corresponding private data is restricted to use in the private large language model. Based on the judgment result, analyze whether the user has violated the data usage regulations. If not, determine the management user of the private data. Calling the private large language model to analyze the online communication records between the user who initiated the data request and the user who manages the private data, and determine whether there is any communication behavior based on the online communication records; Simulate the online communication behavior of the user who initiates the data request and the user who manages the privacy data under compliance conditions, quantify and score the simulated communication behavior, and obtain the first scoring result; Analyze the communication behaviors in the online communication records between the user who initiated the data request and the user who manages the privacy data, and obtain a second scoring result after performing quantitative scoring; Compare the first scoring result with the second scoring result, and determine whether there is any illegal data usage behavior based on the comparison result.

[0015] Furthermore, after determining that a user has engaged in illegal data usage, the user's illegal behavior level is analyzed, specifically including: Determine the upper and lower limits of the communication behavior quantitative scores in the online communication records between the user initiating the data request and the user managing the privacy data, and obtain a score value range consisting of the upper and lower limits; Within the scoring range, simulate the online communication behavior of users who initiate data requests and users who manage privacy data in compliance with regulations. Obtain randomly generated data containing quantitative scores of various communication behaviors, and construct a covariance matrix based on the randomly generated data. Calculating the Mahalanobis distance based on the first scoring result, the second scoring result, and the covariance matrix; Calculate the difference between the second scoring result and the upper and lower scoring limits to obtain the upper and lower scoring limit differences; The difference between the upper and lower limits of the score is summed, the calculated Mahalanobis distance is divided by the sum, and the user's violation level is determined based on the division result.

[0016] Furthermore, the method further includes the following operations: After the task is completed, the AI agent reads the task execution results and performs data verification on the task execution results; Compare the data verification results with the expected targets to determine whether the task execution is successful; If the task is judged to be unsuccessful, the error message is recorded and the task is executed again.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Data security and privacy protection: By deploying a private large language model within the enterprise's internal network, enterprise data can be processed locally, reducing the risk of data leakage. 2. Efficient Customization: By using the company's own data to fine-tune the locally deployed large language model, the model can be better adapted to specific business needs, thereby improving the accuracy and efficiency of task processing; 3. Cost optimization: This invention can reduce an enterprise's reliance on public models, lower call costs, and fully utilize the enterprise's internal computing resources; 4. Intelligent decision support: By combining AI agents with large language models, it can achieve automated processing of complex tasks and intelligent decision-making, improving enterprise operational efficiency; 5. Dynamic resource optimization: AI agents can dynamically allocate resources based on task priority and complexity, improving overall system performance; 6. Autonomous learning and optimization: AI agents can automatically adjust decisions based on historical data and optimize task processing logic to adapt to changes in corporate business. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a schematic diagram of the overall process of an application method based on a privatized large model and an AI agent provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The principles and features of the present invention are described below with reference to the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0021] Reference Figure 1 This embodiment provides an application method based on a private large model and an AI agent, the method comprising: S101. Deploy a pre-trained large language model on the enterprise's internal server, adjust the large language model using the enterprise's own data, and deploy it locally to obtain a private large language model.

[0022] In this step, you first need to select a model architecture that suits the company's needs, then use the company's internal labeled data to fine-tune the large language model, optimize the model's performance in specific fields, and deploy the model to the company's internal high-performance computing cluster to ensure the stability and efficiency of the model's operation.

[0023] S102. Build an AI agent framework to define the AI agent's task execution logic, decision-making rules, and the interactive interface for data interaction with the private large language model.

[0024] In this step, the AI agent framework includes a perception module, a decision module, and an execution module. By defining an interface between the AI agent and the large language model, uninterrupted communication between the two is ensured. The AI agent's decision rules are also initialized, allowing it to dynamically adjust its execution strategy based on task requirements.

[0025] S103. The AI agent receives the task request made by the user in natural language, calls the private large language model to analyze the task request, and generates a task execution plan.

[0026] S104. The AI agent executes the task according to the output of the private large language model and feeds back the task execution results to the user.

[0027] In this embodiment, the AI agent monitors the task execution process, collects execution results and feedback information, and then optimizes the task processing logic based on user feedback after feeding the execution results back to the user. At the same time, the overall efficiency of the system is optimized by regularly evaluating the performance of the private large language model and the AI agent.

[0028] As a possible implementation, before deploying a pre-trained large language model on an internal enterprise server, the following operations are performed: S201. Obtain enterprise demand information, where the enterprise demand information includes business demand information, technical resource information, and data compliance requirement information.

[0029] Exemplarily, the business requirement information includes but is not limited to core application scenarios, performance requirements, and professional field adaptation information. The core application scenarios determine the different requirements of different tasks for model capabilities. Performance requirements include response delay requirements and concurrent processing capability requirements. Professional field adaptation information is used to characterize whether industry-specific models are needed, such as vertical fields such as medical, legal, and financial. Technical resource information includes but is not limited to enterprise hardware conditions, technical team capabilities, existing IT infrastructure, etc. Data compliance requirement information includes but is not limited to data offline deployment requirements, data privacy compliance requirements, knowledge update frequency, etc.

[0030] S202: Input the enterprise demand information into a preset machine learning model for processing to obtain recommended model architecture information.

[0031] S203. Determine the large language model architecture for subsequent deployment based on the recommended model architecture information.

[0032] This implementation method obtains recommended model architecture information by inputting enterprise demand information into a pre-trained preset machine learning model for processing. Enterprises can quickly determine the large language model architecture suitable for themselves based on the recommended model architecture information, thereby speeding up the efficiency of early model deployment.

[0033] As another possible implementation, the AI agent receives a task request from a user in natural language, uses a large language model to analyze the task request, and generates a task execution plan, specifically including: S301: Receive a task request submitted by a user through a natural language interface.

[0034] S302 calls the private large language model to parse the task request, identify the key requirements and goals of the task request, and generate a task to be executed.

[0035] S303: Decompose the task to be executed into multiple subtasks and generate a task execution plan.

[0036] S304: Allocate computing resources according to the task execution plan to complete the execution of the subtask.

[0037] In this implementation, the AI agent receives task requests submitted by users in natural language through a natural language interface. Users can enter task requests in a simple, descriptive manner. After receiving the task request, the AI agent uses a private large language model to parse the task request, identifying the key requirements and objectives of the task request. This understanding of the user's needs generates a task to be executed. By breaking the task into multiple subtasks, the agent generates a corresponding task execution plan and allocates computing resources to achieve efficient parallel processing of the subtasks, improving task processing efficiency.

[0038] On this basis, the task to be executed is decomposed into multiple subtasks to generate a task execution plan, which specifically includes: S401: Analyze the available computing resource information of all computer devices in the enterprise's internal network to obtain an available computing resource evaluation result.

[0039] S042. Decompose the task to be executed according to the evaluation result of available computing resources to obtain a subtask set including multiple subtasks of unequal sizes.

[0040] S403: Allocate each subtask in the subtask set to different computer devices in the enterprise internal network according to its own scale to obtain an allocation result.

[0041] S404: Generate a task execution plan based on the allocation result.

[0042] In this implementation, available computing resource information of all computer devices in the enterprise's internal network is analyzed to obtain an available computing resource evaluation result, and the task to be executed is decomposed according to the available computing resource evaluation result to obtain multiple subtasks of unequal sizes. These subtasks of unequal sizes are respectively allocated to corresponding computer devices for processing according to the size of their required computing resources, thereby making full use of the enterprise's own computing resources to achieve efficient processing of tasks.

[0043] Exemplarily, analyzing the available computing resource information of all computer devices in the enterprise's internal network to obtain an available computing resource evaluation result specifically includes the following operations: S501. Determine the number of computer devices in the enterprise's internal network, as well as a sampling method and computing resource indicators. The sampling method is used to limit the collection time period and collection frequency of the computing resource indicators.

[0044] Exemplarily, the computing resource indicators include but are not limited to CPU usage, GPU usage, memory usage, bandwidth occupancy, etc.

[0045] S502: Collect computing resource indicators of all computer devices according to a sampling method, and generate a computing resource matrix reflecting the real-time computing resource usage of the computer devices.

[0046] In this step, the computing resource indicators of each computer device in the enterprise internal network are collected regularly according to the collection time period specified by the sampling method and the preset collection frequency, and a computing resource matrix is generated based on the three dimensions of computer device information, collection time and computing resource indicators.

[0047] S503: Input the computing resource matrix into the neural network model for processing to obtain the prediction result of the available computing resource information of all computer devices in the enterprise internal network in the future period of time as the available computing resource evaluation result.

[0048] Exemplarily, the neural network model may adopt a pre-trained long short-term memory network, or a neural network model with other structures according to actual needs, which is not specifically limited in this embodiment.

[0049] This implementation method collects computing resource indicators of computer devices in the enterprise's internal network and uses a neural network model to predict the available computing resources of each computer device at a certain moment in the future, so as to facilitate the subsequent allocation of subtasks of different sizes, so that the idle computing resources in the enterprise's internal network can be further fully utilized.

[0050] As another possible implementation, during the execution of the subtasks, the execution order of each subtask is dynamically adjusted. The dynamic adjustment specifically includes regularly performing the following operations: S601: Determine execution ranking indicators of subtasks and normalize the execution ranking indicators.

[0051] In this step, execution ranking indicators include, but are not limited to, importance, urgency, deadline, and task size. These indicators reflect the importance or urgency of the subtasks and are used to guide their execution ranking. The execution ranking indicators are normalized, specifically by converting them to values between [0, 1] to maintain consistency across different types of execution ranking indicators, facilitating subsequent calculation and processing.

[0052] S602: Define a fuzzy set of execution ranking indicators, and further define a membership function of the fuzzy set.

[0053] In this step, a fuzzy set is constructed for each execution ranking metric, and a membership function for the constructed fuzzy set is further defined. The membership function assigns a membership value to each element in the universe, ranging from 0 to 1, reflecting the degree to which the element belongs to the fuzzy set.

[0054] S603 sets a subtask execution sorting rule, applies the subtask execution sorting rule to the fuzzy set, and outputs a sorted fuzzy set.

[0055] Exemplarily, the execution sorting rules may be: important or urgent tasks are sorted first, and relatively unimportant or non-urgent tasks are sorted last; tasks with relatively less remaining completion time are sorted first, and tasks with relatively more remaining completion time are sorted last, etc.

[0056] S604: Defuzzify the sorting fuzzy set to obtain specific execution sorting information of each subtask.

[0057] S605: Adjust the execution order of the subtasks according to the execution order information of each subtask.

[0058] This implementation method first determines the execution ranking index of each subtask separately and normalizes it, then defines the fuzzy set of the execution ranking index and the corresponding membership function, then sets the subtask execution ranking rule, applies the subtask execution ranking rule to the fuzzy set, outputs the ranking fuzzy set, defuzzifies the ranking fuzzy set to obtain accurate execution ranking information, and finally adjusts the execution order of each subtask according to the guidance of the execution ranking information. By reasonably scheduling the execution ranking of each subtask, efficient task processing is achieved, and the occurrence of queuing delays is avoided.

[0059] By deploying the private large language model locally, the risk of data leakage when using public models is reduced to a certain extent. However, daily business operations generate a large amount of data of varying types, each with varying confidentiality levels and the potential impact of a leak. When using a private large language model, the input data may also be used to train the model, affecting responses to users who do not originally have access to the data, potentially leading to data leaks such as salary, financial, and sales data.

[0060] To address the above technical issues, as another possible implementation method, when analyzing task requests, the AI agent uses a private large language model to determine whether the task request contains content that may lead to the leakage of corporate privacy information and to hold relevant users accountable. Specifically, the following operations are performed: S701. The AI agent identifies the identity of the user who initiates the task request and obtains the user identity information, which includes at least the name and position, and determines the user data permissions based on the user identity information.

[0061] S702. Call the privatized large language model to analyze whether there is any possible confidential content in the task request, determine the data type of the possible confidential content, and send the data type of the confidential data to the privacy proof module. The privacy proof module determines the privacy data features used for proof based on the data type of the possible confidential data.

[0062] In this implementation, the privacy proof module is used to verify whether a piece of data is private. However, the AI agent and the privacy proof module do not directly interact with the content of potentially confidential or private data. After obtaining the data type of potentially confidential data, the privacy proof module determines the corresponding data features of the private data based on the data type to verify the potentially confidential data.

[0063] S703, the AI agent and privacy proof module verify whether the possible confidential content is private data based on the characteristics of the possible confidential content and private data through a zero-knowledge proof algorithm. If the possible confidential content is not private data, the task execution plan continues to be generated.

[0064] If the potentially confidential content is private data, the privacy proof module determines the data permissions required to use the private data, and combines the user's data permissions to determine whether the user has violated the data usage regulations. If so, an accountability alarm message is generated.

[0065] In this implementation, the AI agent and the privacy proof module use a zero-knowledge proof algorithm to verify whether potentially confidential content is private data. This eliminates the need for direct interaction between the two involving the actual content of the private data, thus preventing the leakage of private data during interaction. If the potentially confidential content is private data, the privacy proof module further determines whether the user has the authority to legally use the corresponding private data. If not, the user is deemed to have engaged in illegal data usage. Based on the user information and the illegal data usage, an accountability alarm is generated and sent to management personnel to prevent the leakage of corporate private data.

[0066] Exemplarily, the privacy proof module determines the data permissions required to use private data and, based on the user's data permissions, determines whether the user has engaged in illegal data usage, specifically including: S801. Determine whether the user's data permissions meet the data permissions required to use private data. If so, further determine whether the corresponding private data is restricted to use in the privatized large language model. Based on the judgment result, analyze whether the user has violated the data usage regulations. If not, determine the management user of the private data.

[0067] In this step, if the user's data permissions do not meet the data permissions required to use private data, the user is considered to have violated data usage regulations. If the user's data permissions meet the data permissions required to use private data, but the private data they use is restricted to use on a private large language model, the user is also considered to have violated data usage regulations. If the user's data permissions do not meet the data permissions required to use private data, the user who manages the private data is further identified to facilitate data leakage tracing.

[0068] S802: Call the private large language model to analyze the online communication records between the user who initiated the data request and the user who manages the privacy data, and determine whether there is any communication behavior based on the online communication records.

[0069] S803: Simulate the online communication behavior of the user who initiates the data request and the user who manages the privacy data under compliance conditions, perform quantitative scoring on the simulated communication behavior, and obtain a first scoring result.

[0070] S804: Analyze the communication behaviors in the online communication records between the user who initiated the data request and the user who manages the privacy data, and obtain a second scoring result after performing quantitative scoring.

[0071] S805: Compare the first scoring result with the second scoring result, and determine whether there is any illegal data usage behavior based on the comparison result.

[0072] This implementation method, after determining that a user has engaged in illegal data usage, if the user has the authority to access private data, the responsibility for the leakage lies primarily with the user himself. If the user does not have the authority to access private data, it is necessary to determine the administrator of the private data to further determine whether the privacy data leakage is related to the administrator. This implementation method quantifies and scores the simulated online communication behavior and actual communication behavior of the user who initiated the task request and the privacy data administrator under compliance conditions, and calculates the difference between the two to determine whether there is illegal data usage, thereby providing a new solution for privacy data leakage and security protection during the use of privatized large models. If there is no online communication record between the user who initiated the data request and the administrator who manages the private data, it can be temporarily assumed that the privacy data leakage may not be related to the administrator.

[0073] As another further possible implementation, after determining that a user has engaged in illegal data usage, the user's illegal behavior level is analyzed, specifically including: S901: Determine the upper and lower limits of the communication behavior quantitative scores in the online communication records between the user initiating the data request and the user managing the privacy data, and obtain a score value range consisting of the upper and lower limits.

[0074] S902. Within the scoring value range, simulate the online communication behavior of the user who initiates the data request and the user who manages the privacy data under compliance conditions, obtain randomly generated data containing quantitative scores of various communication behaviors, and construct a covariance matrix based on the randomly generated data.

[0075] S903: Calculate the Mahalanobis distance based on the first scoring result, the second scoring result, and the covariance matrix.

[0076] S904: Calculate the difference between the second scoring result and the upper and lower scoring limits, respectively, to obtain an upper and lower scoring limit difference.

[0077] S905: Sum the difference between the upper and lower limits of the score, divide the calculated Mahalanobis distance by the sum, and determine the user's violation level based on the division result.

[0078] This implementation method, based on determining whether a user has engaged in illegal data usage, further simulates the online communication behavior of data users and data managers, constructs a covariance matrix after quantifying the scores, and combines the Mahalanobis distance to determine the user's violation level, thereby more intuitively reflecting the severity of the user's violation, and providing data support for subsequent tracing of relevant responsibilities.

[0079] As another possible implementation, the method further includes the following operations: S1001. After the task is completed, the AI agent reads the task execution result and performs data verification on the task execution result.

[0080] S1002. Compare the data verification result with the expected target to determine whether the task execution is successful; if it is determined that the task execution is unsuccessful, record the error information and re-execute the task.

[0081] This implementation method ensures the correctness and reliability of task execution by adding a data verification mechanism during task execution.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An application method based on a private large model and an AI agent, characterized in that: The method comprises: Deploy a pre-trained large language model on the company's internal server, adjust the large language model using the company's own data, and deploy it locally to obtain a private large language model. Build an AI agent framework, define the AI agent's task execution logic, decision-making rules, and interaction interface with the private large language model; The AI agent receives task requests from users in natural language, uses a private large language model to analyze the task requests, and generates a task execution plan; The AI agent performs tasks based on the output of the private large language model and feeds back the task execution results to the user.

2. The application method based on a private large model and an AI agent according to claim 1, characterized in that: Before deploying a pre-trained large language model on an internal server, perform the following operations: Obtaining enterprise demand information, including business demand information, technical resource information, and data compliance requirement information; Input enterprise demand information into the preset machine learning model for processing to obtain recommended model architecture information; Determine the large language model architecture for subsequent deployment based on the recommended model architecture information.

3. The application method based on a private large model and an AI agent according to claim 1, characterized in that: The AI agent receives task requests from users in natural language, uses the large language model to analyze the task requests, and generates a task execution plan, including: Receive task requests submitted by users through a natural language interface; Call the private large language model to parse the task request, identify the key requirements and goals of the task request, and generate the task to be executed; Decompose the task to be executed into multiple subtasks and generate a task execution plan; Allocate computing resources according to the task execution plan to complete the execution of subtasks.

4. The application method based on a private large model and an AI agent according to claim 3, characterized in that: Decomposing the task to be executed into multiple subtasks and generating a task execution plan specifically includes: Analyze the available computing resource information of all computer devices in the enterprise internal network and obtain the available computing resource evaluation results; Decomposing the task to be executed according to the evaluation result of available computing resources to obtain a subtask set including multiple subtasks of unequal sizes; Allocate each subtask in the subtask set to different computer devices in the enterprise internal network according to its own scale to obtain the allocation result; Generate a task execution plan based on the allocation results.

5. The application method based on a private large model and an AI agent according to claim 4, characterized in that: Analyze the available computing resource information of all computer devices in the enterprise internal network and obtain the available computing resource evaluation results, including: Determine the number of computer devices in the enterprise's internal network, as well as a sampling method and computing resource indicators, wherein the sampling method is used to limit the collection time period and collection frequency of computing resource indicators; Collect computing resource indicators of all computer devices according to a sampling method, and generate a computing resource matrix reflecting the real-time computing resource usage of the computer devices; The computing resource matrix is input into the neural network model for processing to obtain the prediction results of the available computing resource information of all computer devices in the enterprise internal network in the future period of time as the available computing resource evaluation results.

6. The application method based on a private large model and an AI agent according to claim 3, characterized in that: During the execution of subtasks, the execution order of each subtask is dynamically adjusted. Specifically, the following operations are performed regularly: Determine the execution ranking indicators of the subtasks and normalize the execution ranking indicators; Define the fuzzy set of execution ranking indicators and further define the membership function of the fuzzy set; Set the subtask execution sorting rules, apply the subtask execution sorting rules to the fuzzy set, and output the sorted fuzzy set; Defuzzify the sorting fuzzy set to obtain the specific execution sorting information of each subtask; Adjust the execution order of the subtasks according to the execution sorting information of each subtask.

7. The application method based on a private large model and an AI agent according to claim 1, characterized in that: When analyzing a task request, the AI agent uses a private large language model to determine whether the task request contains content that may lead to the leakage of corporate privacy information and to hold the relevant users accountable. This includes the following operations: The AI agent identifies the user who initiated the task request, obtains user identity information, which includes at least name and position, and determines user data permissions based on the user identity information; Call the privatized large language model to analyze whether there is any possible confidential content in the task request, determine the data type of the possible confidential content, and send the data type of the confidential data to the privacy proof module. The privacy proof module determines the privacy data features used for proof based on the data type of the possible confidential data; The AI agent and privacy proof module use a zero-knowledge proof algorithm to verify whether the potentially confidential content is private data based on the characteristics of the potentially confidential content and private data. If the potentially confidential content is not private data, the task execution plan is generated. If the potentially confidential content is private data, the privacy proof module determines the data permissions required to use the private data, and combines the user's data permissions to determine whether the user has violated the data usage regulations. If so, an accountability alarm message is generated.

8. The application method based on a private large model and an AI agent according to claim 7, characterized in that: The privacy proof module determines the data permissions required to use private data and, based on the user's data permissions, determines whether the user has engaged in illegal data usage. Specifically, the module: Determine whether the user's data permissions meet the data permissions required to use private data. If so, further determine whether the corresponding private data is restricted to use in the private large language model. Based on the judgment result, analyze whether the user has violated the data usage regulations. If not, determine the administrator of the private data. Calling the private large language model to analyze the online communication records between the user who initiated the data request and the user who manages the private data, and determine whether there is any communication behavior based on the online communication records; Simulate the online communication behavior of the user who initiates the data request and the user who manages the privacy data under compliance conditions, quantify and score the simulated communication behavior, and obtain the first scoring result; Analyze the communication behaviors in the online communication records between the user who initiated the data request and the user who manages the privacy data, and obtain a second scoring result after performing quantitative scoring; Compare the first scoring result with the second scoring result, and determine whether there is any illegal data usage behavior based on the comparison result.

9. The application method based on a private large model and an AI agent according to claim 8, characterized in that: After determining that a user has engaged in illegal data usage, the user's violation level is analyzed, specifically including: Determine the upper and lower limits of the communication behavior quantitative scores in the online communication records between the user initiating the data request and the user managing the privacy data, and obtain a score value range consisting of the upper and lower limits; Within the scoring range, simulate the online communication behavior of users who initiate data requests and users who manage privacy data in compliance with regulations. Obtain randomly generated data containing quantitative scores of various communication behaviors, and construct a covariance matrix based on the randomly generated data. Calculating the Mahalanobis distance based on the first scoring result, the second scoring result, and the covariance matrix; Calculate the difference between the second scoring result and the upper and lower scoring limits to obtain the upper and lower scoring limit differences; The difference between the upper and lower limits of the score is summed, the calculated Mahalanobis distance is divided by the sum, and the user's violation level is determined based on the division result.

10. The application method based on a private large model and an AI agent according to claim 1, characterized in that: The method further comprises the following operations: After the task is completed, the AI agent reads the task execution results and performs data verification on the task execution results; Compare the data verification results with the expected targets to determine whether the task execution is successful; If the task is judged to be unsuccessful, the error message is recorded and the task is executed again.

Images