Intelligent discriminant optimization system and method based on large language model

Abstract

The invention provides an intelligent discriminant optimization system and method based on a large language model, and the system comprises an intelligent agent management module which is used for managing various types of intelligent agents, and comprises an intelligent discriminant and a plurality of business intelligent agents; the big language model calling module is used for processing different business prompt cues according to the indication of the intelligent discriminant and interacting with corresponding big language models; and the large language model management module comprises three types of models, namely a small parameter model, a large parameter model and a vertical domain model. According to the method, the most suitable large language model is selected for processing, each sub-task is ensured to be processed by the most skilled model, and the accuracy of a processing result is improved.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence and natural language processing, and in particular, to an intelligent agent optimization system and method based on a large language model. Background Art

[0002] In the field of artificial intelligence and natural language processing, especially the development of large language models (LLMs), significant technological progress has been achieved. Large language models, such as GPT-3 and GPT-4, can perform complex tasks such as semantic understanding, text generation and logical reasoning by virtue of their large parameter magnitude and powerful computing power. These models have a wide range of applications, including but not limited to automatic question-answering systems, text summary generation, language translation and dialogue systems. Prior art "Enterprise process processing method and device based on large model" (application number: CN 118229103 A, Beijing Hanbo Technology Co., Ltd.) discloses an enterprise process processing method, which aims to obtain the enterprise process task goal input by the user, rely on the target company intelligent body, select a suitable large model from the model library for reasoning, and predict the task processing method, and finally use the target tool to perform the task to obtain the result, so as to realize the intelligence and automation of the enterprise process. However, the prior art does not solve the bottleneck problem caused by the inability of a single model to efficiently process complex tasks, and relies on a fixed enterprise process model.

[0003] The core of large language models lies in their deep neural networks trained with large amounts of data, which are able to capture the complex patterns and structures of language. Existing large language models usually rely on prompt words to guide the model to generate the desired output. Prompt words are input text designed to stimulate the model to generate a specific type of response. For example, a user can enter a question or a text, and the model generates the corresponding answer or continues the text by understanding and processing these inputs.

[0004] In the application of large models, prompt words play a vital role. A well-designed prompt can significantly improve the output quality of the model. However, the design and optimization of prompts are not easy and require repeated fine-tuning and testing. In the prior art, the design of prompt words mainly relies on manual experience and repeated experiments, and lacks systematic and automated optimization methods.

[0005] Although large language models perform well in many tasks, their application still faces some challenges and limitations. First, the performance of the model is limited by its training data and parameter magnitude. Although GPT-3 has 175 billion parameters and GPT-4 has a higher parameter magnitude, the model's capabilities do not increase linearly and often have a "ceiling" effect. More powerful models require more computing resources and data support, which is costly and difficult to popularize.

[0006] Secondly, when dealing with complex tasks, a single large model often has difficulty understanding and executing complex prompts. This is because large models are designed for generality, and their performance in specific fields or tasks may not be as good as that of specially trained small models. Therefore, solutions that rely on a single model may not achieve the expected results in practical applications.

[0007] In order to solve the above problems, people have tried many methods. For example, by fine-tuning large models to make them perform better on specific tasks, or by using RAG (Retrieval-Augmented Generation) technology to combine the retrieval system and the generative model to improve the accuracy of the model's answers. However, these methods still rely on the capabilities of a single model and cannot completely solve the bottleneck of complex task processing. Summary of the invention

[0008] In view of the shortcomings of the existing technology, this application proposes a general intelligent discriminant solution based on a large language model and prompt words. By introducing the "intelligent discriminant", a set of complex prompt words is split into multiple "business prompt words + business intelligent bodies". The core function of the intelligent discriminant is to intelligently select and allocate suitable models to handle different subtasks according to the complexity and requirements of the task.

[0009] The first aspect of the present application provides an intelligent discriminant optimization system based on a large language model, comprising:

[0010] Agent management module: used to manage various types of agents, including intelligent discriminant agents and multiple business agents;

[0011] Large language model calling module: processes different business prompt words according to the instructions of the intelligent discriminator, and interacts with the corresponding large language model;

[0012] Large language model management module: includes three types of models: small parameter model, large parameter model and vertical field model.

[0013] In combination with the first aspect, the above-mentioned intelligent discriminant is used to receive the initial prompt word, split the task, and dynamically select the business intelligent body and the most suitable model to perform task allocation and integration.

[0014] In combination with the first aspect, the intelligent discriminator further includes a semantic analysis submodule and a task identification submodule:

[0015] The semantic analysis submodule uses a pre-trained language model to perform semantic analysis on the initial prompt word to extract the main task information;

[0016] The task identification submodule identifies multiple subtasks in the initial prompt word through a pre-trained language model and custom rules.

[0017] In combination with the first aspect, the intelligent discriminator uses the following submodules to dynamically select the most suitable model for processing:

[0018] Multi-dimensional feature scoring submodule: scores the features of each business prompt, including task complexity, domain expertise, and response speed requirements;

[0019] Model prioritization submodule: prioritizes models according to their characteristics in the model library;

[0020] The fitness calculation submodule calculates the fitness score of each model and the business prompt, and selects the model with the highest score for processing;

[0021] Real-time feedback adjustment submodule: During task execution, the model selection is dynamically adjusted based on real-time feedback to ensure the optimal processing effect.

[0022] In combination with the first aspect, the above-mentioned business intelligent body processes the business prompt words assigned by the intelligent discriminant body, executes corresponding business processing tasks according to the business prompt words and the selected large language model, and returns the processing results to the intelligent discriminant body.

[0023] Combined with the first aspect, the above-mentioned business intelligence agent uses dependency syntactic analysis technology to conduct in-depth analysis of business prompt words and identify keywords and concepts; calls the knowledge retrieval system and the reverse index-based search algorithm to quickly retrieve information in related fields, and integrates and understands this information through the knowledge graph; uses the Transformer-based text generation model to generate answers based on task characteristics and retrieval results.

[0024] Combined with the first aspect, the above-mentioned large language model calling module is monitored through large model resources, and the resource usage of the server where each large language model is located is monitored in real time through nvitop, including GPU, CPU, and memory; the large model resources are analyzed, the resource monitoring results are analyzed, and different model resources are classified into idle resources and busy resources; the model resources are allocated, and the model resources are optimized according to the resource analysis results.

[0025] In combination with the first aspect, the large language model calling module further includes the following functions:

[0026] Optimize the allocation of model resources based on the results of resource analysis; receive business prompt words and business intelligent entities, receive business prompt words and business intelligent entities from the intelligent discriminant entity; receive model type requests from the intelligent discriminant entity; search for idle resources based on the model type; allocate the idle model resources to business prompt words and business intelligent entities to ensure that tasks can be processed efficiently.

[0027] The second aspect of the present application provides an intelligent discriminant optimization method based on a large language model, comprising:

[0028] Agent management step: used to manage various types of agents, including intelligent discriminant agents and multiple business agents;

[0029] Large language model calling step: processing different business prompt words according to the instructions of the intelligent discriminator, and interacting with the corresponding large language model;

[0030] Large language model management steps: including three types of models: small parameter models, large parameter models and vertical field models.

[0031] A third aspect of the present application provides a computer-readable storage medium storing one or more programs, characterized in that:

[0032] When the one or more programs are executed, an intelligent discriminant optimization method based on a large language model can be implemented.

[0033] Compared with the prior art, this application has the following advantages:

[0034] In terms of task processing efficiency, the intelligent discriminator can split complex tasks into multiple subtasks and dynamically select the most suitable model for processing, thus improving the efficiency of task processing;

[0035] In terms of the accuracy of processing results, the intelligent discriminator can select the most suitable large language model for processing according to the characteristics and requirements of the task, ensuring that each subtask is processed by the most proficient model, thus improving the accuracy of the processing results;

[0036] In terms of resource utilization, the large language model calling module optimizes resource allocation by real-time monitoring and analysis of model resource usage to ensure the efficiency and stability of system operation.

[0037] In terms of scope of application, this application is applicable to various types of tasks and fields, including but not limited to question-answering systems, text generation, machine translation, etc.; in terms of user experience, through the collaborative work of intelligent discriminants and large language models, users can obtain fast and accurate task processing results, thereby improving user experience.

[0038] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures indicated in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 The system structure diagram of the intelligent discriminant optimization system based on the large language model is shown;

[0041] Figure 2 The intelligent discriminant implementation flow chart of the intelligent discriminant optimization system based on the large language model is shown;

[0042] Figure 3 The implementation flow chart of the intelligent discriminant optimization method based on a large language model is shown. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0044] Before describing the specific implementation methods in detail, some special terms mentioned in the technical solution are explained to better understand the technical principles and implementation methods of the present application.

[0045] Large Language Model (LLM)

[0046] Concept: A large language model is a natural language processing model based on deep learning technology, which usually has hundreds of millions to hundreds of billions of parameters and can understand and generate natural language text.

[0047] Function: In this patent, the large language model is used to process various business prompt words and provide high-quality natural language understanding and generation capabilities.

[0048] Technical principle: The large language model learns the statistical laws and semantic features of the language through large-scale corpus training, and can generate coherent and logical text in a given context.

[0049] prompt word

[0050] Concept: The prompt word is the natural language text input by the user, which is used to guide the large language model to generate the corresponding output.

[0051] Function: In this patent, the prompt word is the input for the intelligent discriminant to perform task analysis and splitting.

[0052] Technical principle: The prompt word interacts with the large language model, triggering the model to generate text output related to the prompt word.

[0053] NLP

[0054] Concept: Natural language processing (NLP) is a machine learning technique that enables computers to interpret, process, and understand human language.

[0055] Function: In this patent, NLP is used to classify and extract text, as well as to process the tasks after the splitting.

[0056] Technical principle: When processing natural language, NLP needs to first convert the text into a series of symbols, then analyze and process these symbols, and finally convert these processing results into a language that the machine can understand and execute.

[0057] nvitop

[0058] Concept: It is an open source project developed by Xuehai Pan to provide real-time, detailed system resource monitoring and performance analysis tools for NVIDIA GPU users.

[0059] Function: In this patent, nvitop is used to monitor key indicators such as utilization, memory usage, and computing load of each GPU in real time.

[0060] Technical principle: nvitop is written in Python and uses NVIDIA's nvprof and nvidia-smi command-line tools to obtain the underlying information of the GPU.

[0061] ARIMA

[0062] ARIMA model (English: Autoregressive Integrated Moving Average model), also known as the integrated moving average autoregressive model (moving can also be called sliding), is one of the time series forecasting and analysis methods. In ARIMA (p, d, q), AR is autoregressive, p is the number of autoregressive terms; MA is moving average, q is the number of sliding average terms, and d is the number of differences (order) made to make it a stationary series. Although the word "difference" does not appear in the English name of ARIMA, it is a key step to make the time series stable.

[0063] Ant Colony Optimization Algorithm

[0064] Ant Colony Optimization (ACO), also known as the ant algorithm, is a probabilistic algorithm used to find the optimal path in a graph. It was proposed by Marco Dorigo in his doctoral thesis "Ant system: optimization by a colony of cooperating agents" in 1992, and its inspiration comes from the behavior of ants in finding paths in the process of searching for food. The ant colony algorithm is a simulated evolutionary algorithm. Preliminary studies have shown that the algorithm has many excellent properties. For the PID controller parameter optimization design problem, the results of the ant colony algorithm design are compared with the results of the genetic algorithm design. The numerical simulation results show that the ant colony algorithm has the effectiveness and application value of a new simulated evolutionary optimization method.

[0065] Particle Swarm Optimization

[0066] Particle Swarm Optimization (PSO), also known as particle swarm algorithm, is an evolutionary computing technology developed by J. Kennedy and RCEberhart in 1995, which comes from the simulation of a simplified social model. The "swarm" comes from the particle swarm, which conforms to the five basic principles of swarm intelligence proposed by MMMillonas when developing models for artificial life. "Particle" is a compromise choice, because it is necessary to describe the members of the group as having no mass and no volume, and also to describe its speed and acceleration state.

[0067] After introducing the relevant noun parsing, the implementation process of the general intelligent discriminant solution based on the large language model and prompt words will be described in detail below. This application splits complex tasks into multiple subtasks through intelligent discriminants and dynamically selects the most suitable model for processing. The following are the specific steps and methods.

[0068] This application proposes a general intelligent discriminant solution based on a large language model and prompt words. By introducing an "intelligent discriminant", a set of complex prompt words is split into multiple "business prompt words + business intelligent bodies". The core function of the intelligent discriminant is to intelligently select and allocate suitable models to handle different subtasks according to the complexity and requirements of the task.

[0069] The introduction of intelligent discriminators can effectively solve the following problems:

[0070] 1. Task splitting and allocation: The intelligent discriminator can split complex tasks into multiple subtasks and select the most suitable model for processing according to the characteristics of each subtask. This method not only improves the efficiency of task processing, but also makes full use of the advantages of different models.

[0071] 2. Model selection and optimization: Through the intelligent discriminator, different models can be dynamically selected and switched, including general large models and specially trained small models, to ensure that each subtask is handled by the best model. This flexibility enables the system to better cope with complex and diverse task requirements.

[0072] 3. Improve processing capabilities: The division of labor mechanism of the intelligent discriminator enables the system to better understand and process complex prompt words, improving the processing capabilities and accuracy of the overall system.

[0073] This application solves the bottleneck problem of a single model processing complex tasks in the prior art by introducing an intelligent discriminator. The intelligent discriminator optimizes the processing flow of prompt words through task splitting and model selection, and improves the overall performance and task processing capabilities of the system. This innovative solution not only improves the application effect of large language models, but also provides new ideas and methods for the development of future intelligent systems.

[0074] like Figure 1 As shown, the technical solution of this application includes the following core modules:

[0075] Intelligent agent management module: used to manage various types of intelligent agents, including intelligent discriminant agents and multiple business intelligent agents.

[0076] Large language model calling module: According to the instructions of the intelligent discriminator, it processes different business prompt words and interacts with the corresponding large language model.

[0077] Large language model management module: includes various types of models, such as small parameter models, large parameter models, and vertical field models.

[0078] The agent management module includes intelligent discriminant agents and business agents, among which the intelligent discriminant agent is the core component of this application, responsible for analyzing and splitting the initial prompt words, and dynamically selecting the most suitable model for processing. Its specific functions include:

[0079] 1. Receive the initial prompt word: The intelligent discriminator receives the initial prompt word input by the user and analyzes the requirements and complexity of the task.

[0080] 2. Task splitting: Through NLP technology, the intelligent discriminator splits complex tasks into multiple business prompts. For example, a complex question-and-answer task can be split into multiple simple question-and-answer subtasks.

[0081] 3. Select business agent: According to the characteristics of the business prompt, the intelligent discriminator selects the corresponding business agent for processing. The business agent can be a general agent, a professional agent or a simple agent.

[0082] 4. Select a suitable model: The intelligent discriminator selects the most suitable large language model for processing according to the characteristics and requirements of the task. Model types include small parameter models, large parameter models and vertical domain models.

[0083] 5. Task allocation and integration: The intelligent discriminator passes the business prompt words and the business agent to the large language model calling module for processing, receives and integrates the processing results, and finally returns the results to the user.

[0084] The business agent is responsible for processing the business prompt words assigned by the intelligent discriminator and interacting with the corresponding large language model. Its specific functions include:

[0085] 1. Receiving business prompt words: The business intelligent body receives the business prompt words transmitted by the intelligent discriminant body.

[0086] 2. Execute business processing: Based on the business prompt words and the selected large language model, the business agent executes the corresponding business processing tasks.

[0087] 3. Return processing results: The business intelligence agent returns the processing results to the intelligent discriminant agent.

[0088] The large language model calling module is responsible for processing different business prompt words according to the instructions of the intelligent discriminator and interacting with the corresponding large language model. Its specific functions include:

[0089] Large model resource monitoring: Use nvitop to monitor the resource usage of the servers where each large language model is located in real time, including GPU, CPU, memory, etc.

[0090] Large model resource analysis: Analyze the resource monitoring results and classify different model resources into idle resources and busy resources.

[0091] Model resource allocation: Based on the resource analysis results, optimize the allocation of model resources, further including:

[0092] Receive business prompt words and business intelligent body: Receive business prompt words and business intelligent body from the intelligent discriminant body.

[0093] Receive model type request: Receive model type request from the intelligent discriminator.

[0094] Find idle resources: Find idle resources based on model type.

[0095] Allocate model resources: Allocate idle model resources to business prompts and business agents to ensure that tasks can be processed efficiently.

[0096] The large language model management module includes multiple types of models, each of which is good at handling different types of tasks. Specifically, they include:

[0097] Small parameter model: suitable for processing simple tasks, with lower computing resource requirements and faster response speed.

[0098] Large parameter models: Suitable for handling common and complex tasks, providing higher accuracy and stronger processing capabilities.

[0099] Vertical field model: specially designed to handle tasks in a specific field and can provide specialized solutions.

[0100] The implementation process of the intelligent discriminant solution is as follows: Figure 2 As shown,

[0101] 1. Implementation process of intelligent discriminator:

[0102] 1.1 Receiving the initial prompt word, the intelligent discriminator first receives the initial prompt word input by the user. This prompt word can be any form of natural language text, such as a question, instruction, or statement. For example, the prompt word input by the user is: "Please explain the basic principles of quantum computing."

[0103] 1.2 Task splitting: The intelligent discriminant uses natural language processing (NLP) technology to analyze the initial prompt and identify the key tasks and subtasks. For example, for the question "Please explain the basic principles of quantum computing", the intelligent discriminant may split it into the following subtasks:

[0104] Explain the definition of quantum computing; describe the concept of quantum bit (qubit); introduce basic quantum computing operations such as quantum gates and quantum entanglement.

[0105] 1.3 Select a business agent. According to the characteristics of each business prompt, the intelligent discriminant selects the most suitable business agent for processing. Business agents can be general agents, professional agents, or simple agents. For example, to explain the definition of quantum computing, the intelligent discriminant may select a general agent; and to describe the concept of quantum bits, it may select a professional agent that specializes in the field of science and technology.

[0106] 1.4 Select a suitable model. The intelligent discriminator selects the most suitable large language model for processing according to the characteristics and requirements of the task. Model types include small parameter models, large parameter models, and vertical domain models. For example, to explain the definition of quantum computing, a small parameter model may be selected; while to describe the concept of quantum bits, a large parameter model or a vertical domain model may be selected.

[0107] 1.5 Task allocation and integration, the intelligent discriminant passes the business prompt and business agent to the large language model call module for processing, receives and integrates the processing results, and finally returns the results to the user. For example, the intelligent discriminant assigns the task of explaining the definition of quantum computing to the general agent, and the task of describing the concept of quantum bits to the professional agent. After calling the corresponding large language model for processing, the results of each subtask are integrated to form the final explanation text and returned to the user.

[0108] 2. Task splitting algorithm and model selection mechanism

[0109] One of the core functions of the intelligent discriminator is task splitting and model selection. This process uses natural language processing (NLP) technology and a series of algorithms to ensure accuracy and efficiency.

[0110] 2.1. Task Splitting Algorithm

[0111] The intelligent discriminator uses a pre-trained language model based on the Transformer architecture (such as BERT or GPT) for initial analysis to identify key tasks and subtasks in the user input prompt. The steps are as follows:

[0112] Semantic analysis: Use the pre-trained language model to perform semantic analysis on the initial prompt words and extract the main task information.

[0113] Task identification: Identify multiple subtasks in the initial prompt using a pre-trained language model and custom rules.

[0114] 2.2. Model selection mechanism

[0115] To select the most suitable model for each subtask, the intelligent discriminator adopts a dynamic selection mechanism, which includes the following steps:

[0116] Task feature extraction: Extract features from each subtask prompt. The features include task type, complexity, domain knowledge requirements, etc.

[0117] Model matching: According to the characteristics of the task, the most suitable model is selected from the model library, which includes small parameter models, large parameter models and vertical field models. The selection criteria are as follows:

[0118] Small parameter model: suitable for simple tasks, fast response speed and low computing resource consumption.

[0119] Large parameter models: suitable for both common and complex tasks, and can provide higher accuracy.

[0120] Vertical field model: specially designed to handle tasks in a specific field and can provide specialized solutions.

[0121] 3. Dynamic model selection algorithms and rules

[0122] The intelligent discriminator uses a comprehensive set of algorithms and rules to dynamically select the most suitable model for processing:

[0123] Multi-dimensional feature scoring: Score the features of each business prompt, including task complexity, domain expertise, response speed requirements, etc.

[0124] Model prioritization: Prioritize models based on their characteristics in the model library (such as number of parameters, training datasets, domain expertise, etc.).

[0125] Fitness calculation: Calculate the fitness score between each model and the business prompt, and select the model with the highest score for processing.

[0126] Real-time feedback adjustment: During task execution, model selection is dynamically adjusted based on real-time feedback (such as response time, resource consumption, etc.) to ensure optimal processing results.

[0127] The intelligent discriminant optimization method based on the large language model of this application is as follows Figure 3 As shown, the following steps are included:

[0128] Agent management step: used to manage various types of agents, including intelligent discriminant agents and multiple business agents;

[0129] Large language model calling step: processing different business prompt words according to the instructions of the intelligent discriminator, and interacting with the corresponding large language model;

[0130] Large language model management steps: including three types of models: small parameter models, large parameter models and vertical field models.

[0131] The specific implementation process of the business intelligence body is as follows:

[0132] 1. Implementation process

[0133] 1.1 Receive business prompt

[0134] The business agent receives the business prompt word transmitted by the intelligent discriminant. For example, the business prompt word received by the business agent may be: "Explain the concept of quantum bit".

[0135] 1.2 Execute business processing

[0136] According to the business prompt and the selected large language model, the business agent performs the corresponding business processing task. For example, the business agent calls the large language model to generate an explanation text about the concept of quantum bits.

[0137] 1.3 Return processing results

[0138] The business agent returns the processing result to the intelligent discriminant. For example, the explanation text generated by the business agent may be: "A quantum bit is the basic unit of quantum computing, which can simultaneously represent the superposition state of 0 and 1."

[0139] 2. Specific technical details

[0140] 2.1 Task Analysis

[0141] Use natural language processing technology, such as dependency analysis, to conduct in-depth analysis of business prompts and identify keywords and concepts.

[0142] 2.2 Knowledge retrieval and integration

[0143] The business agent calls a specially designed knowledge retrieval system and a reverse index-based search algorithm to quickly retrieve information in related fields and integrate and understand this information through a knowledge graph.

[0144] 2.3 Content Generation

[0145] Use a Transformer-based text generation model (such as GPT-x, etc.) to generate answers based on task characteristics and retrieval results.

[0146] The specific implementation process of the large language model calling module is as follows:

[0147] 1. Implementation process

[0148] Large model resource monitoring: nvitop is used to monitor the resource usage of the servers where each large language model is located in real time, including GPU, CPU, memory, etc. For example, the large language model call module can monitor the GPU usage of the server to ensure the efficiency of resource allocation.

[0149] Large model resource analysis: Analyze resource monitoring results and classify different model resources into idle resources and busy resources. For example, if the GPU usage of a server is less than 20%, it is classified as an idle resource.

[0150] Model resource allocation:

[0151] 1.1 Optimize the allocation of model resources based on the results of resource analysis.

[0152] 1.2 Receive business prompts and business agents: Receive business prompts and business agents from the intelligent discriminator.

[0153] 1.3 Receive model type request: Receive the model type request from the intelligent discriminator.

[0154] 1.4 Find idle resources: Find idle resources according to the model type.

[0155] 1.5 Allocate model resources: Allocate idle model resources to business prompts and business agents to ensure that tasks can be processed efficiently. For example, after the large language model call module receives the business prompts and professional agent requests to explain the concept of quantum bits, it searches for idle large parameter model resources and allocates them to the task.

[0156] 2. Specific technical details

[0157] 2.1 Real-time monitoring: Use professional tools (such as nvitop) and custom scripts to monitor the utilization of GPU, CPU and other resources in real time.

[0158] 2.2 Dynamic analysis and resource optimization and allocation: The collected data is analyzed and predicted through time series prediction models (such as ARIMA), and then resources are allocated in combination with load balancing and resource optimization algorithms (such as ant colony optimization and particle swarm optimization).

[0159] 2.3 Model calling and optimization: Use microservice architecture to dynamically schedule different model instances, and use container technology (such as Docker) to isolate and manage the operating environment of each model.

[0160] The large language model management module specifically includes the following modules:

[0161] Small parameter model: Suitable for processing simple tasks, with low computing resource requirements and fast response speed. For example, for explaining the definition of quantum computing, you can choose a small parameter model to quickly generate an explanation text.

[0162] Large parameter model: Suitable for processing common and complex tasks, it can provide higher accuracy and stronger processing power. For example, for describing the concept of quantum bits, a large parameter model can be selected for processing to generate more detailed and accurate explanation text.

[0163] Vertical domain models: They are specialized for processing tasks in a specific field and can provide specialized solutions. For example, for advanced concepts and operations involving quantum computing, vertical domain models can be selected to generate highly specialized explanatory texts.

[0164] This application has the following application scenarios:

[0165] Application scenario 1: Processing questions and answers in the financial field

[0166] 1. The user enters the initial prompt: "Please explain the application of blockchain in the financial field."

[0167] 2. The intelligent discriminator receives and analyzes the initial prompt words and splits the task into multiple business prompt words: "Definition of blockchain", "Technical principles of blockchain", "Application of blockchain in the financial field".

[0168] 3. The intelligent discriminator selects the appropriate business agent and large language model, and passes the business prompt to the large language model calling module.

[0169] Select a general agent to process "Blockchain Definition" and use a small parameter model.

[0170] Select professional agents to process "Technical principles of blockchain" and use large parameter models.

[0171] Select the vertical domain agent to process "Application of blockchain in the financial field" and use the vertical domain model.

[0172] 4. The large language model calling module allocates appropriate model resources to process business prompt words based on resource monitoring and analysis results.

[0173] Allocate the resources of server B to vertical domain agents to process "the application of blockchain in the financial field".

[0174] 5. The business intelligence agent performs business processing tasks and returns the results to the intelligent discriminant agent.

[0175] The general agent returns "Blockchain is a distributed ledger technology."

[0176] The professional intelligent agent returned that “the technical principles of blockchain include decentralization, consensus mechanism and encryption technology.”

[0177] The vertical field agent returned "The applications of blockchain in the financial field include cross-border payments, smart contracts and supply chain finance."

[0178] 6. The intelligent discriminator integrates the processing results of each subtask and returns the final result to the user. Application scenario 2: Processing medical question answering

[0179] 1. The user enters the initial prompt: "Please explain the application of gene editing technology in medicine."

[0180] 2. The intelligent discriminator receives and analyzes the initial prompt words and splits the task into multiple business prompt words: "Definition of gene editing", "Technical principles of gene editing", "Application of gene editing in medicine".

[0181] 3. The intelligent discriminator selects the appropriate business agent and large language model, and passes the business prompt to the large language model calling module.

[0182] Select the general agent to handle "Definition of Gene Editing" and use a small parameter model.

[0183] Select a professional agent to process the "Technical principles of gene editing" using a large parameter model.

[0184] Select the vertical domain agent to process "Application of gene editing in medicine" and use the vertical domain model.

[0185] 4. The large language model calling module allocates appropriate model resources to process business prompt words based on resource monitoring and analysis results.

[0186] Allocate the resources of server B to the vertical domain agent to process "Application of gene editing in medicine".

[0187] 5. The business intelligence agent performs business processing tasks and returns the results to the intelligent discriminant agent.

[0188] The general agent returns "Gene editing is a technology that modifies genes by changing their DNA sequences."

[0189] Professional intelligence returns "The technical principles of gene editing include CRISPR-Cas9, TALEN and ZFN".

[0190] The vertical field agent returned "The applications of gene editing in medicine include the treatment of genetic diseases, cancer and infectious diseases."

[0191] 6. The intelligent discriminator integrates the processing results of each subtask and returns the final result to the user.

[0192] Application scenario 3: Processing questions and answers in the education field

[0193] 1. The user enters the initial prompt: "Please explain the advantages and challenges of online education."

[0194] 2. The intelligent discriminator receives and analyzes the initial prompt words and splits the task into multiple business prompt words: "Definition of online education", "Advantages of online education", and "Challenges of online education".

[0195] 3. The intelligent discriminator selects the appropriate business agent and large language model, and passes the business prompt to the large language model calling module.

[0196] A general agent was selected to process “Definition of Online Education” using a small parameter model.

[0197] Select professional agents to handle the “advantages of online education” and use large parameter models.

[0198] Select vertical domain agents to deal with the "Challenges of Online Education" and use the vertical domain model.

[0199] 4. The large language model calling module allocates appropriate model resources to process business prompt words based on resource monitoring and analysis results.

[0200] Allocate the resources of server B to vertical domain agents to handle the "challenges of online education".

[0201] 5. The business intelligence agent performs business processing tasks and returns the results to the intelligent discriminant agent.

[0202] The general agent returns "Online education is a form of education conducted through the Internet."

[0203] The professional intelligence body returned that “the advantages of online education include flexibility, rich resources, and personalized learning.”

[0204] Vertical field intelligence returns that "the challenges of online education include technical barriers, insufficient interaction between teachers and students, and difficulty in evaluating learning outcomes."

[0205] 6. The intelligent discriminator integrates the processing results of each subtask and returns the final result to the user.

[0206] Based on the same application concept, the present application also provides a computer-readable storage medium storing one or more programs. When the one or more programs are executed, the aforementioned electricity market transaction decision-making method with the solar thermal power station as the main body can be implemented.

[0207] The present application has the following advantages: in terms of task processing efficiency, complex tasks are split into multiple subtasks through intelligent discriminators, and the most suitable model is dynamically selected for processing, thereby improving the efficiency of task processing; in terms of processing result accuracy, the intelligent discriminator can select the most suitable large language model for processing according to the characteristics and requirements of the task, ensuring that each subtask is processed by the most proficient model, thereby improving the accuracy of the processing results; in terms of resource utilization, the large language model calling module optimizes resource allocation by real-time monitoring and analysis of model resource usage, thereby ensuring the efficiency and stability of system operation; in terms of scope of application, the present patent solution is applicable to various types of tasks and fields, including but not limited to question-answering systems, text generation, machine translation, etc.; in terms of user experience, through the collaborative work of the intelligent discriminator and the large language model, users can obtain fast and accurate task processing results, thereby improving user experience.

[0208] Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that it is still possible to modify the technical solutions described in the aforementioned embodiments, or to make equivalent replacements for some of the technical features therein; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent discriminant optimization system based on a large language model, characterized in that: include: Agent management module: used to manage various types of agents, including intelligent discriminant agents and multiple business agents; Large language model calling module: according to the instructions of the intelligent discriminator, it processes different natural language text prompt words and interacts with the corresponding large language model; Large language model management module: includes three types of models: small parameter model, large parameter model and vertical field model.

2. The system according to claim 1, characterized in that The intelligent discriminator is used to receive the initial prompt word, split the task, and dynamically select the business intelligent body and the most suitable model to perform task allocation and integration.

3. The system according to claim 2, characterized in that The intelligent discriminator further includes a semantic analysis submodule and a task identification submodule: The semantic analysis submodule uses a pre-trained language model to perform semantic analysis on the initial prompt word to extract the main task information; The task identification submodule identifies multiple subtasks in the initial prompt word through a pre-trained language model and custom rules.

4. The system according to claim 3, characterized in that The intelligent discriminator uses the following submodules to dynamically select the most suitable model for processing: Multi-dimensional feature scoring submodule: scores the features of each business prompt, including task complexity, domain expertise, and response speed requirements; Model prioritization submodule: prioritizes models according to their characteristics in the model library; The fitness calculation submodule calculates the fitness score between each model and the business prompt, and selects the model with the highest score for processing; Real-time feedback adjustment submodule: During task execution, the model selection is dynamically adjusted based on real-time feedback to ensure the optimal processing effect.

5. The system according to claim 2, characterized in that The business intelligence body processes the business prompt words assigned by the intelligent discriminator, executes corresponding business processing tasks according to the business prompt words and the selected large language model, and returns the processing results to the intelligent discriminator.

6. The system according to claim 5, characterized in that The business agent uses dependency syntactic analysis technology to deeply analyze business prompt words and identify keywords and concepts; calls the knowledge retrieval system and uses a reverse index-based search algorithm to quickly retrieve information in related fields, and integrates and understands this information through a knowledge graph; uses a Transformer-based text generation model to generate answers based on task characteristics and retrieval results.

7. The system according to claim 1, characterized in that The large language model calling module monitors through large model resources, and monitors the resource usage of each server where the large language model is located in real time through the system resource monitoring and performance analysis tool nvitop, including GPU, CPU, and memory; analyzes the large model resources, analyzes the resource monitoring results, and classifies different model resources into idle resources and busy resources; allocates model resources, and optimizes the allocation of model resources according to the resource analysis results.

8. The system according to claim 7, characterized in that The large language model calling module further includes the following functions: Optimize the allocation of model resources based on the resource analysis results; receive business prompts and business agents, receive business prompts and business agents from intelligent discriminators; receive model type requests from intelligent discriminators; find free resources based on model types; The idle model resources are allocated to business prompt words and business agents to ensure that the tasks can be processed efficiently.

9. A method for optimizing an intelligent discriminant based on a large language model, characterized in that: include: Agent management step: used to manage various types of agents, including intelligent discriminant agents and multiple business agents; Large language model calling step: processing different business prompt words according to the instructions of the intelligent discriminator, and interacting with the corresponding large language model; Large language model management steps: including three types of models: small parameter models, large parameter models and vertical field models.

10. A computer-readable storage medium storing one or more programs, characterized in that: When the one or more programs are executed, the intelligent discriminant optimization method based on a large language model as described in claim 9 can be implemented.

Images