Intelligent agent construction and application system based on large language model

By constructing an intelligent agent system based on a large language model, the problems of high training cost and insufficient generalization ability of large models are solved, realizing efficient and secure intelligent agent applications. It can self-organize, integrate, and schedule multiple model tools to complete complex tasks.

CN118966267AActive Publication Date: 2024-11-15NANJING UNIV OF SCI & TECH

Patent Information

Application Number
CN202411119788.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-15
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The learning process of large models requires a lot of computing resources and time, resulting in high training costs. Furthermore, they lack generalization ability when facing new problems and scenarios, and data security and privacy protection issues are prominent, making it difficult to achieve highly customized intelligent agent applications.

Method used

Construct an intelligent agent building and application system based on a large language model, including an application management module, an intelligent agent workflow module, an intelligent agent building module, and an intent understanding module. Through the large model, understand task requirements, generate code, and combine multiple model tools to achieve self-organizing integration and efficient scheduling of complex tasks.

Benefits of technology

It improves the operational efficiency of intelligent systems, enabling them to accurately understand user intent, self-organize and integrate various applications, generate code to implement complex tasks, reduce training costs, and enhance the flexibility and security of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966267A_ABST
    Figure CN118966267A_ABST
Patent Text Reader

Abstract

The invention discloses an agent construction and application system based on a large language model, which provides a unified registration and release normal form of model service, deeply understands task business requirements through a large model, generates a service arrangement process aiming at user task requirements, and provides a service arrangement process for the user task requirements if the required service does not exist. And if not, driving the code generation capability of the large model to perform service definition and service implementation, continuously calling self iteration in the process, and finally forming an intelligent agent working process for the business. The method has the capabilities of problem semantic understanding and complex task decomposition and calling, and has the advantages that the large model can accurately understand the intention of a user, schedule or compile a plurality of different model tools on an immediate machine and combine to complete the complex task while the large model can be efficiently accessed to an application system, so that the system operation efficiency is greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to intelligent agent construction technology, and more particularly to an intelligent agent construction and application system based on a large language model. Background Art

[0002] With the rapid development of science and technology, artificial intelligence (AI) has become a hot topic in today's world. Among them, large models represented by ChatGPT, as a technology with broad application prospects, are attracting more and more attention from the world.

[0003] A big model is an automated learning system based on artificial intelligence technology. It can learn to extract useful information from a large amount of data and make predictions and decisions based on this information. The development of big models can be divided into two stages: the first stage is the traditional rule-based expert system; the second stage is the automatic learning system based on machine learning. At present, the research on big models is mainly focused on the second stage, that is, the automatic learning system based on deep learning.

[0004] In recent years, deep learning technology has made significant progress, providing strong support for the development of large models. In particular, the emergence of convolutional neural networks (CNN) and recurrent neural networks (RNN) has enabled computers to achieve breakthrough results in image recognition, speech recognition, natural language processing, and other fields. These achievements have provided rich application scenarios for the development of large models, enabling them to achieve remarkable results in various fields.

[0005] In the field of natural language processing, large models can already complete tasks such as text classification, sentiment analysis, and machine translation. For example, Google's BERT model has achieved leading performance in multiple natural language processing tasks. In the field of computer vision, large models can recognize various image information such as faces and objects, providing support for applications such as driverless cars and intelligent monitoring. In addition, large models can also be applied to recommendation systems, games, medical care, and other fields, bringing huge potential for social and economic development.

[0006] However, the development of large models also faces some challenges. First, the learning process of large models requires a lot of computing resources and time, which leads to an increase in its R&D costs. Second, the generalization ability of large models still needs to be improved, that is, when faced with new problems and scenarios, the performance of large models may not be as good as that of human experts. In addition, the data security and privacy protection issues of large models cannot be ignored. How to make full use of the advantages of big data while ensuring data security is an urgent problem to be solved in the development of large models.

[0007] In short, as a technology with broad application prospects, large models have achieved remarkable results in various fields. However, the development of large models still faces many challenges, which require us to continue research and exploration. From the perspective of development trends, the future of large models can be divided into the following aspects:

[0008] (1) Larger model scale: With the improvement of computing power and the accumulation of big data, large models will become larger in the future. Larger model scale means more parameters and richer knowledge, which will enable large models to have higher accuracy and efficiency when processing complex tasks. At the same time, the training of large-scale models also requires more powerful hardware support, so we can foresee that more hardware devices specifically for large models will appear in the future, such as dedicated AI chips.

[0009] (2) Better performance optimization: In order to improve the performance of large models in various application scenarios, researchers will continue to explore more efficient algorithms and optimization methods. For example, techniques such as knowledge distillation and model pruning can reduce computing and storage requirements while maintaining model performance, allowing large models to run on low-power devices. In addition, researchers will also try to use the model's pre-training capabilities to develop more efficient multi-task learning methods to achieve better performance optimization.

[0010] (3) Wider application areas: As big model technology continues to mature, its application areas will continue to expand. In addition to existing scenarios such as intelligent customer service and voice assistants, big models will also play an important role in the fields of medicine, education, finance, etc. For example, in the medical field, big models can assist doctors in disease diagnosis and treatment recommendations; in the education field, big models can provide students with personalized learning suggestions and resource recommendations. These applications will greatly improve people's quality of life and social efficiency.

[0011] (4) Stronger interpretability: Current large models are often viewed as “black boxes” whose internal working principles and decision-making processes are difficult to understand. However, in order for people to better trust and use large models, researchers need to improve the interpretability of models. Through visualization technology, local models and other methods, we can understand the decision-making basis of large models, thereby ensuring that their application in key areas is more reliable and secure.

[0012] (5) Stronger privacy protection: As data leakage and abuse become increasingly serious, how to protect the privacy of user data in large models has become an important issue. Researchers will use technical means such as differential privacy and federated learning to protect the privacy and security of user data. At the same time, corresponding regulations and standards will be introduced at the policy level to ensure that the application of large models in various fields can fully respect the privacy rights and interests of users.

[0013] In short, the big model demonstrates strong universality, creativity, and emergence, providing a new "AI+" paradigm for the intelligent development of the cybersecurity and informatization system, and achieving a leap from "point intelligence" to "system intelligence". Summary of the invention

[0014] The purpose of the present invention is to provide an intelligent agent construction and application system based on a large language model to solve the problems of highly customized intelligent agent applications, high training costs, and time-consuming and labor-intensive manual work. Through the versatility, creativity, and emergence of large models, the system can accurately understand the user's intentions, dispatch multiple different model tools, and combine them to complete complex tasks, acting as a "model commander", thereby greatly improving the efficiency of system operation.

[0015] The technical solution to achieve the purpose of the present invention is: an intelligent agent construction and application system based on a large language model, including an application management module, an intelligent agent workflow module, an intelligent agent construction module and an intention understanding module;

[0016] The application management module is used to register and publish business services, manage service semantic mapping, and monitor and manage the normality of applications;

[0017] The agent workflow module is used to build a service workflow for a certain task requirement. First, based on the existing application list, the task requirement can be fully understood through the large model, and a suitable agent application can be selected from the application list to drive it;

[0018] The agent building module is used to iterate and generate sub-requirements and codes through the thinking chain capability of the large model when the requirements cannot be met, where the code is verified in the form of function definition, function implementation, function integration implementation and execution result testing;

[0019] The intention understanding and execution module is used to parse the user's natural language needs, match them according to the semantic similarity with the needs, and recommend the best agent for execution; if the user confirms, the execution results and the execution tool flow will be reflected in the feedback to the user.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] (1) Aiming at the complex and changeable characteristics of future intelligent scenarios, the present invention enables the intelligent system to have the ability to understand scenario requirements and call various applications. It can use the big model as the core to perform consistency management modeling on various existing applications and then drive them, so that various applications have task-oriented self-organization integration capabilities;

[0022] (2) The present invention utilizes the ability of the large model to generate code on demand, maps task requirements into program code, and can compile and assemble into specific software applications based on the generated code, so that the application software has strong scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is the overall architecture diagram of the intelligent agent construction and application system based on the large language model of the present invention.

[0024] Figure 2 This is a processing flow chart of the present invention during the execution of the intelligent agent workflow stage.

[0025] Figure 3 This is a processing flow chart of the present invention in the stage of building a new intelligent agent.

[0026] Figure 4 The present invention is a processing flow chart of the execution phase based on task intention understanding. DETAILED DESCRIPTION

[0027] In order to solve the problems mentioned in the background technology, the present invention provides an intelligent agent construction and application system based on a large language model, provides a unified registration and release paradigm for model services, deeply understands the task business requirements through the large model, generates a service orchestration process for user task requirements, and if the required service does not exist, drives the code generation capability of the large model to perform service definition and service implementation, and continuously calls self-iteration in the process, and finally forms an intelligent agent workflow for the business. The present invention has the ability to understand problem semantics, decompose and call complex tasks. The advantage is that while being able to efficiently access the application system, the large model can accurately understand the user's intentions, schedule or write multiple different model tools on the spot, and combine them to complete complex tasks, thereby greatly improving the efficiency of system operation.

[0028] The technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0029] Example

[0030] Combined with Figure 1 The overall architecture of an intelligent agent construction and application system based on a large language model disclosed in an embodiment of the present application is described. The overall architecture hierarchy includes: an application management module, an intelligent agent workflow module, an intelligent agent construction module, and an intent understanding module.

[0031] The application management module includes the application publishing registration module, the semantic description mapping module, the load balancing service pool and the application testing tool. The application publishing registration module is used to uniformly register various applications, describe the basic information, scope of use, and interpretable and evolvable requirements of the application, so that the large model can understand the semantic level and then call it; the semantic description mapping module is used to store the semantic description information of the service, and realize the mapping and calling of semantics and applications by calculating the similarity between the existing intelligent agent capability description and the demand; the load balancing service pool is used to store the service name and service address information, and select the appropriate application for calling according to the current state of the resource pool. If all types of applications in the resource pool are under heavy load, instance resources are dynamically allocated through the container cloud by adding application instances; the application testing tool ensures the normality of the registered application logic function by automatically running the application interface of the intelligent agent, and at the same time ensures the normal state of the application through the service heartbeat mechanism, so that it can be called as an application in the workflow.

[0032] The intelligent agent workflow module mainly includes five modules: large model understanding engine, application list management, prompt statement generation management, workflow construction and workflow execution. The large model understanding engine can integrate various basic large language models and understand them according to user needs. The application list management provides a unified display and management tool for various applications, and can provide consistent application management. The prompt statement generation management can provide various prompt statements that trigger the large model capabilities, and can provide supporting tools such as writing, debugging, and searching for prompt statement templates, so that prompt engineers can write prompts. Workflow construction provides the ability to create, edit and store processes run by multiple intelligent agents based on the logical understanding ability of the large model. The workflow execution module can call each intelligent agent module in sequence through service calls and record the results of the calls.

[0033] The agent construction module mainly provides six modules: agent requirement generation, function parameter definition, code generation, code automatic testing, agent storage management, and container group installation tool. Agent requirement generation can convert agent task requirements that cannot be directly mapped into agent construction requirements. The three modules of function parameter definition, code generation, and code automatic testing will automate the planning of the code writing and testing phase in the agent construction process based on the large model, and finally iterate to generate executable code. The agent storage management module solidifies the generated code into an executable program for storage. The container group installation tool packages the executable program and related dependencies into an installation startup script, which can be quickly started during deployment.

[0034] The intention understanding execution module mainly provides four modules: speech conversion tool, candidate agent generation, operation process monitoring, and user feedback management. The language conversion tool can convert the voice information input by the user into text information. Candidate agent generation can select the top-ranked agent from multiple agents for application to call. Operation process monitoring can monitor the execution process of the agent and the user's feedback on the key results of the execution process. The user feedback management module can standardize and organize various feedback behaviors in the user's behavior into records for incremental learning.

[0035] The specific business service functions of the application release registration module include: completing service registration according to business requirements, and providing basic information such as the service name, address, parameter name and value range, capability description, etc.; required data knowledge information, task scenarios, task levels and other usage scope information, as well as interpretability and evolvability requirements, etc.;

[0036] In the massive text analysis business, event extraction from text is an important functional module in text analysis. Since the extracted content needs to be updated regularly according to the characteristics of the text, its applicability is also limited due to the limitations of training data. Each instance of story event extraction registers, publishes and sends information as follows:

[0037]

Basic Information

[0038] Service name: Text event extraction

[0039] Service address: http: / / 192.168.xx.xxx:8888 / eventExtract

[0040] Parameter name: [{parameter name: text to be extracted},{parameter type: string},{parameter value range:

[0041] No more than 100 characters}]

[0042] Ability description: Based on the input text, extract information such as event type, event participants, event trigger words, event time, event location, etc.

[0043] [Scope of use information]

[0044] Required data service information: national data set, location data set, knowledge graphs of various devices

[0045] Applicable article types: more rigorous content such as news released by official Internet media.

[0046] Entities that can be extracted: name, country, place, time, equipment

[0047] Types of events that can be extracted: assemblies, declarations, protests, etc.

[0048] [Explanable and evolvable requirements]

[0049] Explanability requirements: post-result explanation elements, algorithm process explanation elements, knowledge reasoning explanation elements

[0050] Evolvable requirements: operating status evaluation indicator system, retraining strategy (regular, indicator threshold), incremental adjustment strategy, user feedback information demonstration strategy requirements

[0051] The semantic description mapping module is used to store the semantic description information of the service. By calculating the similarity between the existing agent capability description and the requirements, the semantics and application mapping and calling are realized. The similarity algorithm can use multiple indicators such as cosine similarity and string comparison Rouge-L. If the similarity reaches the threshold, it is considered that the requirements are met, and the task can be executed by the agent, and the execution of each node of the agent is carried out in sequence.

[0052] The load balancing service pool is used to store service names and service address information. According to the current state of the resource pool, the appropriate application is selected for calling. If all types of applications in the resource pool are under heavy load, instance resources are dynamically allocated through the container cloud by adding application instances. Specifically, under the premise that the service startup script can be successfully started, when the number of visits per unit time and the service delay reach a certain threshold, the service script is automatically started to add instances.

[0053] The application testing tool ensures the normal function of the registered application logic by automatically running the application interface of the intelligent agent. At the same time, it ensures that the application status is normal and can be called as an application in the workflow through the service heartbeat mechanism.

[0054] Combined with Figure 2 The work of executing the agent workflow stage is explained. The agent workflow module can fully understand the task requirements through the large model based on the existing application list and select the appropriate agent application from the application list to drive it. The specific steps are as follows:

[0055] A. Build a large model understanding engine. The large model can understand the application list and the requirements of the task.

[0056] B. Organize all kinds of applications in the application management module in a unified way, and obtain the basic information, usage scope and other elements of the corresponding applications from the application management module according to different demand types;

[0057] C. The requirements and application lists are combined into prompt statements and input into the big model understanding engine. The big model generates a process for solving the task through heuristic solution.

[0058] D. The large model uses its own logical understanding ability to call multiple applications and feedback the results. When the feedback results meet the user's needs, the calling process can be solidified and saved as a fixed flowchart. If the feedback results are not satisfactory, they can be manually edited to customize the sequence relationship and input and output;

[0059] E. Finally, according to the process execution task, the execution status of each node in the execution process can be monitored. If all are executed successfully, the intelligent agent workflow can be saved; if the execution fails and iteration is required, return to step D; if the execution is successful and no iteration is required, save it.

[0060] Wherein step A comprises:

[0061] A-1. Incremental learning of the large model. The data for the learning part is more than 10,000 service call data based on user commands. The specific data format can be as follows: {"appName":"Remote sensing image target recognition","ability":"Ability to identify various types of targets appearing in remote sensing images and identify models"};

[0062] A-2. After the incremental training forms a model, the model can be fine-tuned through prompt learning. The fine-tuning uses 10,000 data. The specific data format can be as follows: {"requireMent": "Determine whether the location is a military airport", "appName": "Remote sensing image target recognition", "reason": "It can be determined by judging whether the equipment model on the airport is a military aircraft"};

[0063] A-3. After training the model, encapsulate the function into a callable Web service through Python.

[0064] Wherein step B comprises:

[0065] B-1. Construct field matching rules for the application management module based on user task requirements, which can generate different fields for different types of requirements. Specific judgments can be inferred through production rules. The following example shows that the content extracted for military entity recognition requirements needs to include weapons and equipment:

[0066] if (requirement.type like military) then scope of use. entity contains weapons and equipment

[0067] B-2. Based on different user needs, different application content fields are mapped and assembled to form a comprehensive description statement of the application, such as the following:

[0068] Service name: Text event extraction

[0069] Parameters: [{parameter name: text to be extracted}, {parameter type: string}, {parameter value range: no more than 100 characters}]

[0070] Ability description: Based on the input text, extract information such as event type, event participants, event trigger words, event time, event location, etc.

[0071] Entities that can be extracted: name, country, place, time, weapons and equipment

[0072] Wherein step C comprises:

[0073] C-1. Based on the candidate application list, user demand input, small sample data, output data format, etc., a prompt statement is formed and input into the big model. The big model can return the calling program planning path in the corresponding format;

[0074] C-2. The program automatically parses the content of the program planning path to confirm whether all the applications of the planning path can be called. If it cannot be called or the application does not exist, it uses the multi-round dialogue capability of the large model to iteratively ask questions until a callable program planning path is fully formed or the iteration limit has been reached.

[0075] Combined with Figure 3 The work of building a new agent is explained. The agent building module can unify the data, knowledge and other elements involved in the agent workflow module to ensure independent execution, and then install and deploy based on the container. Some nodes in the workflow need to be implemented under the premise of human-computer interaction because there is no corresponding service implementation. It is mainly divided into the following four steps:

[0076] A. Generate the node sub-requirements that the agent lacks, taking the required functions in the process as input and converting them into data format requirements and code function descriptions;

[0077] B. Generate specific function names and parameter definitions through the thinking chain capability of the big model; Generate specific function implementations based on the specific function names and parameter definitions through the code generation capability of the big model, and compile and run the code through the compiler;

[0078] C. Perform internal testing of the module using the test tool (unittest / JUnit). If the test is successful, proceed to step D, otherwise return to step B.

[0079] D. Finally, the agent conducts a trial run through the generated process. If there is no error, the node process is stored in the workflow. If there is an error, it returns to process B for iteration.

[0080] Wherein step B comprises:

[0081] B-1. Generate two types of requirements, namely, specific data flow and program control flow, based on the node sub-requirements that the intelligent agent lacks. The data flow includes the input and output formats of the program, which are expressed in the form of JSON key-value; the program control flow is mainly provided in the form of code. Based on the current large model training expectations, python and java languages ​​are more common, mainly generating the definition content of the function and the corresponding input and output;

[0082] B-2. Generate function implementations corresponding to function definitions based on the required content of nodes in the program control flow, and implement program logic through multiple languages ​​such as Java and Python;

[0083] B-3. ​​Iterate all nodes in the workflow to implement all function definitions and function functions in the entire process;

[0084] B-4. Execute the workflow to achieve unified convergence of results and solidify the capabilities of the intelligent agent after manual judgment.

[0085] Combined with Figure 4 The execution phase based on task intention understanding is explained. The intention understanding execution module analyzes the user's natural language requirements, selects similar processes from the perspectives of semantic similarity and context, and generates specific intelligent agent execution results through combined calls. The specific steps are as follows.

[0086] A. Recognize and analyze user voice and convert it into text requirements described in natural language. For specific implementation examples, you can use tools for converting user requirements into text that the model can understand.

[0087] B. Associate the text requirements with the capabilities that the intelligent agent can handle, mainly by performing comprehensive calculations from multiple dimensions such as the semantic similarity of the literal text, the context in which the requirements are raised, and the referential relationship. Based on the similarity threshold, a candidate set that can handle the requirements is screened out. Specifically, the cosine distance can be used for calculation. Based on the cosine distance value, a candidate set that can handle the requirements is screened out, and a trial run is performed after the user makes the selection.

[0088] C. Monitor the input and output of various nodes in the process in real time, and clearly display the operation status logs of all process nodes and the specific contents of input and output, which is convenient for real-time monitoring.

[0089] D. Finally, the content will be fed back to the user and the running process will be stored.

[0090] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An intelligent agent construction and application system based on a large language model, characterized in that: It includes application management module, intelligent agent workflow module, intelligent agent construction module and intention understanding module; The application management module is used to register and publish business services, manage service semantic mapping, and monitor and manage the normality of applications; The agent workflow module is used to build a service workflow for task requirements. It can fully understand the task requirements through a large model based on the existing application list and select a suitable agent application from the application list to drive it; The agent building module is used to iterate and generate sub-requirements and codes through the thinking chain capability of the large model when the requirements cannot be met, where the code is verified in the form of function definition, function implementation, function integration implementation and execution result testing; The intention understanding and execution module is used to parse the user's natural language needs, match them according to the semantic similarity with the needs, and recommend the best agent for execution; if the user confirms, the execution results and the execution tool flow will be reflected in the feedback to the user.

2. The intelligent agent construction and application system based on a large language model according to claim 1 is characterized in that: The application management module includes an application publishing and registration module, a semantic description mapping module, a load balancing service pool, and an application testing tool; the application publishing and registration module is used to uniformly register various applications, describe the basic information, usage scope, and interpretable and evolvable requirements of the application, so that the large model can understand the semantic level and then call it; The semantic description mapping module is used to store the semantic description information of the service, and realizes the mapping and calling of semantics and application by calculating the similarity between the requirements; the load balancing service pool is used to store the service name and service address information, and selects the appropriate application for calling according to the current state of the resource pool. If the load of various applications in the resource pool exceeds the set threshold, they are dynamically allocated through the container cloud by adding application instances; the application testing tool ensures the normal status of the registered application by automatically running the application interface of the intelligent body, so that it can be called as an application in the workflow.

3. The intelligent agent construction and application system based on a large language model according to claim 1, characterized in that: The agent workflow module can fully understand the task requirements through the large model based on the existing application list, and select the appropriate agent application from the application list to drive it. The specific implementation is as follows: A. Build a large model understanding engine. The large model can understand the application list and the requirements of the task. B. Organize all kinds of applications in the application management module in a unified way, and obtain the basic information and usage scope of the corresponding applications from the application management module according to different demand types; C. The requirements and application lists are combined into prompt statements and input into the big model understanding engine. The big model generates a process for solving the task through heuristic solution. D. The large model uses its own logical understanding ability to call multiple applications and feedback the results. When the feedback results meet the user's needs, the calling process is solidified to form a fixed flow chart for storage. When the feedback results do not meet the user's needs, manual editing is performed to customize the sequence relationship and input and output. E. According to the process execution task, the execution status of each node in the execution process can be monitored. If all are executed successfully, the agent workflow can be saved; If the execution fails and iteration is required, return to step D; if the execution succeeds and iteration is not required, save the data.

4. The intelligent agent construction and application system based on a large language model according to claim 3 is characterized in that: The specific process of building a large model understanding engine includes: Incremental learning of large models, where the learning data is service call data based on user commands; After the model is formed through incremental training, the model is fine-tuned through prompt learning; After the model is trained, the functions are encapsulated into a callable Web service through Python.

5. The intelligent agent construction and application system based on a large language model according to claim 3 is characterized in that: The various applications of the application management module are organized in a unified manner, and according to different types of requirements, the basic information and usage scope of the corresponding applications can be obtained from the application management module, specifically including: Build application management module field matching rules based on user task requirements, and generate different query fields for different requirement types; Different application content fields are mapped based on different user needs and assembled to form a comprehensive description statement of the application.

6. The intelligent agent construction and application system based on a large language model according to claim 3 is characterized in that: The process of combining the requirements and application lists into prompt statements and inputting them into the big model understanding engine, and the big model generating a solution to the task through heuristic solution, includes: Based on the candidate application list, user demand input, small sample data, and output data format, prompt statements are formed and input into the big model. The big model returns the calling program planning path in the corresponding format. The program automatically parses the content of the calling program planning path to confirm whether all the applications of the planning path can be called. If it cannot be called or the application does not exist, it uses the large model's multi-round dialogue capability to iteratively ask questions until a callable program planning path is fully formed or the iteration limit has been reached.

7. The intelligent agent construction and application system based on a large language model according to claim 1 is characterized in that: The agent building module can unify and package the data and knowledge elements involved in the agent workflow module to ensure independent execution, and then install and deploy it based on the container; the nodes that cannot be successfully mapped in the workflow need to be implemented under the premise of human-computer interaction because there is no corresponding service implementation, which mainly includes: A. Generate the node sub-requirements that the agent lacks, taking the required functions in the process as input and converting them into data format requirements and code function descriptions; B. Generate specific function names and parameter definitions through the thinking chain capability of the big model; Generate specific function implementations based on the specific function names and parameter definitions through the code generation capability of the big model, and compile and run the code through the compiler; C. Perform internal testing of the module through the test tool. If it runs smoothly, proceed to step D, otherwise return to step B; D. Finally, the agent conducts a trial run through the generated process. If there is no error, the node process is stored in the workflow. If there is an error, it returns to process B for iteration.

8. The intelligent agent construction and application system based on a large language model according to claim 7 is characterized in that: The aforementioned method uses the thinking chain capability of the big model to generate specific function names and parameter definitions; uses the code generation capability of the big model to generate specific function implementations based on the specific function names and parameter definitions, and compiles and runs the code through the compiler, specifically including: According to the node sub-requirements that the intelligent agent lacks, two types of requirements are generated: specific data flow and program control flow. The data flow includes the input and output formats of the program, which are expressed in the form of JSON key-value. The program control flow is mainly provided in the form of code, which mainly generates the definition content of the function and the corresponding input and output. Generate the function implementation corresponding to the function definition according to the required content of the nodes in the program control flow, and implement the program logic through Java and Python multi-language; Iterate all nodes in the workflow to implement all function definitions and function functions in the entire process; Execute the workflow to achieve unified convergence of results, and solidify the capabilities of the intelligent agent after manual judgment.

9. The intelligent agent construction and application system based on a large language model according to claim 1, characterized in that: The intention understanding and execution module analyzes the user's natural language needs, selects similar processes from the perspective of semantic similarity and context, and generates specific agent execution results through combined calls. The specific implementation method is as follows: A. Recognize and analyze user voice and convert it into textual requirements described in natural language; B. Associate the text requirements with the capabilities that the agent can handle, mainly by performing comprehensive calculations from multiple dimensions such as the semantic similarity of the text, the context in which the requirements are raised, and the referential relationship. Select a candidate set that can handle the requirements based on the similarity threshold, and run the test after the user selects it. C. Monitor the input and output of various nodes in the process in real time, and clarify the operation status of all process nodes and the specific content of input and output, which is convenient for real-time monitoring; D. Finally, the content will be fed back to the user and the running process will be stored.

Citation Information

Patent Citations

  • Intelligent dialogue method and system based on natural language processing

    CN118132719A

  • Context-aware and persona-based generative ai interpretation device and method for controlling the same

    KR102692549B1

  • Ubiquitous operating system oriented toward group intelligence perception

    WO2021213293A1

Cited By

  • Water conservancy business dynamic response method and system based on large language model

    CN119294561A

  • Adaptive large model AI application construction system based on multi-model logic arrangement

    CN119415078A

  • Vehicle-mounted user intention recognition and task generation system based on large model

    CN119416895A

  • Vehicle-mounted user intent recognition and task generation system based on large model

    CN119416895B

  • Intelligent agent construction method and system based on large language model, medium and equipment

    CN119512523A