AI native application construction platform system and method based on pre-trained large model

Abstract

The invention discloses an AI native application construction platform system and method based on a pre-trained large model. The system comprises a data processing flow module which provides a pipeline for collecting data resources to data assets; the model processing flow module is used for providing model development and training support for the artificial intelligence model; the asset management module is used for providing various assets and asset management services and providing on-demand access through various interface protocols; the multi-agent development assembly line module comprises an agent cluster and realizes agent cluster management, task decomposition and execution of the agent cluster, task feedback and operation monitoring; and the intelligent agent cluster is supported to be embedded into an enterprise business system, so that interaction between multiple intelligent agents and a business process is realized. According to the method, a multi-agent cluster can be driven, a large model is called to execute adaptive application assembly in combination with a specific scene, and a solution for constructing an AI native application is provided in a complex scene in the vertical field.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI native application building platform system and method based on pre-trained large models. Background Technology

[0002] Industrial applications are mostly based on classic artificial intelligence capabilities, such as deep learning methods based on regression or classification. These are only suitable for discriminative and well-defined single-point scenarios. The introduction of large-scale pre-trained models has greatly propelled the evolution of artificial intelligence from standalone applications to more general AI capable of adapting to diverse scenarios. This has facilitated greater exposure to and use of artificial intelligence by ordinary end-users.

[0003] Large Language Models (LLMs) based on natural language have demonstrated significantly improved performance compared to traditional methods in multiple scenarios, but limitations still exist in industrial applications. Unlike general large language modeling tasks, industrial scenarios in vertical fields are characterized by low fault tolerance, special features, and strong dependence on professional experience, which makes it difficult for users to stimulate the emergent capabilities of large language models through simple human-computer interaction.

[0004] Furthermore, various enterprise information systems are driven by data, processes, mechanisms, and principles, with all modules interconnected to form a complex system. The software interfaces adapted to each system can differ, and the challenges in building and assembling intelligent applications introduce uncertainty into the implementation of large-scale language model applications. Therefore, a platform is needed that can adapt to the interfaces of various enterprise information systems.

[0005] The application of large-scale language models in industry involves various professional fields such as exploration, production, sales, safety control, financial cost control, and human resource management. Alignment gaps exist between the business language of end users and the technical language of developers, posing numerous challenges beyond technical aspects to system design and application development. Furthermore, supporting multiple language model-based applications requires sufficient high-quality data assets, including data based on business mechanisms and process-generated data. This data typically possesses private domain attributes and multimodal properties. These data assets are wide-ranging, numerous, and inconsistent, necessitating comprehensive data storage, governance, transmission, representation, and application mechanisms to ensure their reliability. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide an AI native application construction platform system and method based on pre-trained large models that overcomes or at least partially solves the above problems.

[0007] In a first aspect, embodiments of the present invention provide an AI native application building platform system based on a pre-trained large model, comprising:

[0008] The data processing workflow module provides a complete pipeline from data resource collection to data asset identification, supporting alignment, feature fusion, content fusion, and tag fusion of multiple modal data; it also supports toolchain capabilities for data asset processing.

[0009] The model processing module is used to provide support for the development and training of artificial intelligence models;

[0010] The asset management module provides various assets and asset management services, and offers on-demand access through multiple interface protocols.

[0011] The multi-agent development pipeline module includes an agent cluster containing multiple agents, and implements agent cluster management, task decomposition and execution, task feedback and operation monitoring to form a closed-loop link that supports the full lifecycle management of the agent cluster; and supports embedding the agent cluster into enterprise business systems to realize the interaction between multiple agents and business processes.

[0012] In one embodiment, the asset management module specifically includes: a data asset module, an artificial intelligence model asset module, a business model module, and an application model management service module, and provides on-demand access to existing information systems through Restful, SOAP, websocket, and GraphQL protocols.

[0013] In one embodiment, the data processing module is used to support data synchronization from heterogeneous data sources such as Mini / O, object storage, MySQL database, SQL Server, NiFi, and vector database; and to support alignment, feature fusion, content fusion, and tag fusion operations for structured or unstructured multimodal data such as speech, images, natural language, and tables.

[0014] In one embodiment, the data processing module specifically includes: a visualization label management component and a data mining and development module, which supports multiple data mining frameworks and is compatible with multiple data annotation and development languages, for realizing the fusion and extraction of multimodal data features.

[0015] In one embodiment, the model processing module and the model training module support artificial intelligence frameworks such as TensorFlow, PyTorch, PaddlePaddle, and Mindsphere, provide data annotation and have a visual interface to support training, evaluation, and debugging, and publish the model through its own inference platform after training is completed.

[0016] In one embodiment, the agents in the multi-agent development pipeline module are used to receive user requests issued in natural language, perform machine understanding on the user intent corresponding to the requests, break down the task chain based on the understood user intent, form corresponding AI native applications, execute the work plan of the AI ​​native applications and drive execution and application debugging. During the debugging process, they achieve link integration and semantic alignment with existing enterprise business systems. After successful joint debugging of multiple agents, an interface is generated, embedded into the existing enterprise business system, executes the work plan corresponding to the requests, and returns the execution results to the user.

[0017] In one embodiment, the joint debugging of multiple agents is achieved through self-supervised reflection and continuous optimization within the multiple agents.

[0018] In one embodiment, multiple agents also achieve adaptive agent retraining and optimization through user feedback on execution results.

[0019] In one embodiment, the system has the capability to construct AI-native applications for planning, recommendation, search, scheduling, prediction, classification, recognition, judgment, dialogue, intent decomposition and application assembly, and multi-agent collaboration.

[0020] Secondly, embodiments of the present invention provide a method for building AI native applications, which builds AI native applications through an AI native application building platform system based on pre-trained large models as described above.

[0021] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0022] The AI ​​native application building platform system and method based on pre-trained large models provided in this invention can drive multi-type and multi-modal intelligent agent clusters. Combined with specific scenarios, it calls large language models to perform adaptive assembly. In complex scenarios in vertical domains, it provides a solution for building AI native applications, performing training or fine-tuning of intelligent agents and large language models, testing and verification, and concise deployment. It also achieves agile integration with existing enterprise business information systems. Compared with traditional agile development and waterfall development, the AI ​​native application building platform system and method based on pre-trained large models provided in this invention significantly improves upon traditional agile development and waterfall development methods in terms of application selection accuracy, application assembly efficiency, application iteration efficiency, large language model selection and scheduling, and the effectiveness of multi-agent management.

[0023] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0026] Figure 1 This is an architecture diagram of the AI ​​native application building platform system based on a pre-trained large model in an embodiment of the present invention;

[0027] Figure 2 and Figure 3 This is a flowchart illustrating how user intent is broken down, task chains are generated, work plans are formed, AI native applications are created, and embedded into existing business systems in this embodiment of the invention.

[0028] Figure 4 This is a schematic diagram of the task of selecting and dismantling the potential layer for re-evaluation of old tight gas wells in an embodiment of the present invention;

[0029] Figure 5 This is a flowchart illustrating the specific implementation of the AI ​​native application for old well re-examination in this invention.

[0030] Figure 6 This is a schematic diagram of a specific example of an AI native application construction platform system based on a pre-trained large model, as described in this invention. Detailed Implementation

[0031] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0032] The inventors of this application have discovered that, due to the lack of support from high-quality data assets in specialized vertical domains, pre-trained large models suffer from insufficient capabilities in reasoning, judgment, interpretation, plan decomposition, and decision support within those vertical domains. Furthermore, vertical domains are characterized by low fault tolerance, specificity, and strong reliance on specialized experience.

[0033] At present, it is difficult to meet the needs of users facing complex scenarios by directly applying general-purpose large models or by fine-tuning them using specialized data. These complex scenario needs include, but are not limited to:

[0034] (1) Intelligent applications struggle to achieve agile assembly: Modern industrial production processes are highly automated, with the entire process chain tightly interconnected. Enterprise business information systems, including Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), Supply Chain Management (SCM), Supplier Relationship Management (SRM), Customer Relationship Management (CRM), and Human Resource Management (HRM), are primarily process-driven, supplemented by data and model-driven approaches. Currently, Large Language Models (LLMs) and their native applications are detachable and replaceable external attachments to enterprise business systems. Relying solely on LLMs to provide services makes it difficult to integrate them effectively with existing business systems and processes.

[0035] (2) High-quality data acquisition chain: The application of LLM requires sufficient high-quality data assets. These data assets are characterized by wide coverage, large quantity and strong inconsistency. This requires complete data storage, governance, transmission, representation and application links to ensure the reliability of the entire data asset chain.

[0036] (3) Alignment of interdisciplinary terminology: The application of large-scale models in vertical industries involves multiple different disciplines such as exploration, production, sales, safety control, financial cost control, and human resource management. The knowledge systems and value logics of different disciplines are not entirely the same. Complex systems with interdisciplinary languages, professional knowledge, and business models urgently need a unified alignment medium to bridge the gap between the business language of end users and the technical language of developers.

[0037] (4) Poor interpretability of the conclusions: From a technical point of view, AIGC still adopts a black box model based on the encoding-decoding architecture, which leads to the risk of creating illusions and providing incorrect guidance in the data services provided by AIGC. It is necessary to have business-related interpretations of the results to meet the understanding and usage needs of end users.

[0038] Based on the above requirements, in order to realize the construction of AI native applications in vertical fields, this embodiment of the invention provides an AI native application construction platform system based on pre-trained large models. This platform system can integrate intelligent agents, industrial data and large language models to simplify the service construction and assembly process based on artificial intelligence generated content.

[0039] To facilitate understanding, this application first provides a brief explanation of the concept of AI-native applications.

[0040] AI-native applications originate from the reconstruction or creation of various application scenarios and their solutions using generative AI concepts (creating new application architectures). For example, in human-computer interaction, products are based on natural language interaction to generate personalized results that adapt to users; in terms of data and information, massive amounts of unstructured data can be directly stored, used, and retrieved, rather than heavily relying on structured data; and in terms of infrastructure, a hardware and software architecture that can meet the requirements of running large models is built.

[0041] Existing AI-native applications possess the following characteristics: They are based on natural language interaction, where users interact with the backend through a language user interface (LUI), requiring little or no interaction with a graphical user interface (GUI), ultimately presenting a hybrid GUI and LUI interaction format. This allows users to leap from limited input to unlimited input, providing both high-frequency, fixed functions and the ability to understand and handle low-frequency, customized needs; they possess autonomous learning and adaptability, integrating understanding, memorization, and adaptation to multimodal data during human-computer interaction, and self-learning, enabling more accurate and personalized adjustments to output results based on changes in context, task environment, and interaction objects; and they possess the ability to autonomously complete tasks, capable of executing precise tasks based on large language models and knowledge bases, achieving an end-to-end closed loop, integrating the entire process from task acquisition to task completion.

[0042] The AI ​​native application building platform system based on pre-trained large models provided in this embodiment of the invention refers to... Figure 1 As shown, it includes:

[0043] The data processing workflow module provides a complete pipeline from data resource collection to data asset identification, supporting alignment, feature fusion, content fusion, and tag fusion of multiple modal data; it also supports toolchain capabilities for data asset processing.

[0044] The model processing module is used to provide support for the development and training of artificial intelligence models;

[0045] The asset management module provides various assets and asset management services, and offers on-demand access through multiple interface protocols.

[0046] The multi-agent development pipeline module includes an agent cluster containing multiple agents, and implements agent cluster management, task decomposition and execution, task feedback and operation monitoring to form a closed-loop link that supports the full lifecycle management of the agent cluster; and supports embedding the agent cluster into the enterprise business system to realize the interaction between multiple agents and business processes.

[0047] In one embodiment, the aforementioned asset management module specifically includes: a data asset module, an artificial intelligence model asset module, a business model module, and an application model management service module. It provides on-demand access to existing information systems through RESTful (a software architecture based on the HTTP protocol for creating, designing, and providing web services), SOAP (Simple Object Access Protocol, a protocol for communication between applications based on XML and HTTP), WebSocket (a network protocol for full-duplex communication over a single TCP connection, such as for communication between browsers and servers), and GraphQL (a protocol about how to communicate with APIs for API calls).

[0048] In one embodiment, the above data processing module can support data synchronization from heterogeneous data sources such as MinI / O, object storage, MySQL database, SQL Server, NiFi (an open-source data component), and vector database; and can support alignment, feature fusion, content fusion, and tag fusion operations for structured or unstructured multimodal data such as speech, images, natural language, and tables.

[0049] The data processing workflow module includes a built-in visual tag management component, data mining, and development modules. It is compatible with common data mining frameworks such as TensorFlow (an open-source machine learning framework), PyTorch (an open-source Python machine learning library), PaddlePaddle (an open-source deep learning platform), and MindSphere (a cloud-based open IoT operating system), and is also compatible with development languages ​​such as C++, Python, R, and MATLAB. It can support and provide a complete toolchain for data asset processing.

[0050] In one embodiment, the aforementioned model processing module is designed based on an independent architecture for training and inference to support the separation of training and inference requirements in industrial application scenarios. It supports mainstream frameworks such as TensorFlow, PyTorch, PaddlePaddle, and Mindsphere, provides data annotation, and has a visual interface to support training, evaluation, and debugging. After model training is complete, it is deployed through its own inference platform.

[0051] In one embodiment, the agents in the aforementioned multi-agent development pipeline module are specifically used to receive user requests issued through natural language, perform machine understanding on the user intent corresponding to the requests, decompose the task chain based on the understood user intent, construct the corresponding AI native application, execute the work plan of the AI ​​native application and drive execution and application debugging, and during the debugging process, achieve link integration and semantic alignment with existing enterprise business systems, and after successful joint debugging of multiple agents, generate an interface, embed it into the existing enterprise business system, execute the work plan corresponding to the requests, and return the execution results to the user.

[0052] In this embodiment of the invention, the task chain (containing several sub-tasks) obtained after decomposing the user intent forms a work plan based on the AI ​​native application. Each sub-task in the work plan can be implemented independently by a single intelligent agent, or more often, each sub-task can be the result of collaborative work by multiple intelligent agents. That is, the execution result of a sub-task undertaken by one intelligent agent may be a necessary condition for the sub-tasks undertaken by other intelligent agents, or the execution result of a sub-task undertaken by one intelligent agent may affect the sub-tasks undertaken by other intelligent agents. In these cases, the intelligent agents must be jointly debugged to avoid misinterpretation of the user intent and to ensure the correctness of the overall understanding of the user intent. Only after successful debugging can the interface be generated and the business system embedded.

[0053] In one embodiment, refer to Figure 2 and Figure 3 The flowchart shown includes the following steps:

[0054] I. User intent is to interact with the intelligent agent (AI-AGENT, or AGENT for short) through natural language to achieve machine understanding;

[0055] Second, the intelligent agent breaks down the task chain based on the understood user intent, forms a work plan, and drives its execution (AI native application generation) and application debugging.

[0056] Third, during the debugging process, the AI ​​native application building platform system based on pre-trained large models provided in this embodiment of the invention can achieve functions such as linking and integrating with existing systems and semantic alignment.

[0057] Fourth, after completing the joint debugging, generate interfaces to perform cross-modal heterogeneous system integration and embed it into existing enterprise business systems.

[0058] The purpose of joint testing is to examine whether the intelligent agent can correctly understand the user's intent, break down the task chain, and, based on the joint testing results, ensure that the user's intent is correctly understood and that the broken down task chain is correct before embedding it into various enterprise business systems (ERP, MES, etc.). This can effectively avoid the impact of incorrect intent understanding on the normal processes of existing enterprise business systems.

[0059] For example, in the application scenario of re-evaluating old logging wells, the user can enter the request "Please display the knowledge graph of the re-evaluation of old logging wells" in the human-computer interaction dialog box of the large model. Then the intelligent agent will retrieve the data of the corresponding knowledge graph and display it on the human-computer interaction interface.

[0060] For example, when a user enters "Please provide the workflow for selecting the potential layer of old tight gas logging wells for re-evaluation" in the human-computer interaction dialog box, the AI ​​will break down the workflow for obtaining the potential layer of old tight gas logging wells for re-evaluation according to the user's intention. The AI ​​will then break it down into the following workflow and feed it back to the human-computer interaction interface of the large model.

[0061] The main steps in the re-evaluation of potential layers for old tight gas wells are as follows:

[0062] Map-driven data preparation → Well logging curve standardization → Intelligent well logging curve reconstruction → Intelligent well logging micro-layer segmentation → Intelligent reservoir identification → Intelligent well logging parameter calculation → Intelligent fluid identification → Intelligent potential layer selection…

[0063] It also displays a schematic diagram of the corresponding workflow breakdown, which can be found in [reference]. Figure 4 As shown.

[0064] Figure 4 The flowchart shown includes the following steps:

[0065] Curve standardization, curve reconstruction, well logging micro-layer segmentation, lithology identification, reservoir identification, well deviation correction, parameter calculation, data classification, fluid identification, production capacity classification, production capacity prediction, potential layer selection, and measure layer recommendation, etc.

[0066] After correctly understanding the user's intent regarding the "workflow for selecting the potential layer for re-evaluation of old tight gas logging wells," the task chain for "selecting the potential layer for re-evaluation of old tight gas wells" is broken down as described above.

[0067] Again Figure 5 The flowchart shown illustrates the process of building an AI-native application.

[0068] Based on the AI ​​native application building platform system provided in the embodiments of the present invention, agile interaction between business systems, large language models and application scenarios can be realized, thereby achieving flexible construction, continuous optimization and agile embedding of large model native applications with existing enterprise information systems.

[0069] like Figure 5 As shown in the embodiment of the old well re-inspection, through the AI ​​native application building platform system provided by the present invention, users (including professional developers or non-professional developers such as prompting engineers) can describe the application requirements of the scenario based on natural language. The platform system utilizes natural semantic understanding capabilities and a self-supervised reflection mechanism to generate the construction code of the corresponding intelligent agent cluster, realizing semantic understanding, task decomposition, application assembly, application integration, task execution, and continuous reflection for user needs; at the same time, it is based on self-supervised reflection and continuous optimization within the intelligent agent and between intelligent agent clusters.

[0070] The AI ​​native application building platform system based on pre-trained large models provides a closed-loop mechanism for multi-agent clusters based on LLM. It can achieve accurate and efficient application selection, application assembly and debugging, large model selection, AI atomic capability scheduling, continuous prompting, agent cluster operation orchestration, business system integration, and adaptive agent retraining and optimization based on user feedback in the loop.

[0071] Figure 5 In the process shown, the user inputs statements regarding the re-evaluation of old wells in the human-computer interaction interface. Based on the pre-trained large model, the AI ​​native application in the platform system constructs a pre-trained large model to understand the intent and decompose the task chain, and specify the execution work plan. These two steps are completed by the intelligent agent. Then, the task chain is executed, and the well logging curve is reconstructed (some existing data may be incomplete or inaccurate, this step relies on the AI ​​model). The curve is then interpreted (this step calls CAD, CAE, other business systems, and knowledge graphs). Potential layer recommendations are made (this step calls AI models and knowledge graphs). Then, the data collection strategy is generated (relying on AI models, knowledge graphs, etc.). Finally, system documents are generated in the enterprise business system (such as ERP system), and the approval decision process and the post-evaluation process are executed.

[0072] Figure 6 This is a specific example of an AI native application building platform system based on a pre-trained large model provided in this embodiment of the invention. In this example, the AI ​​native application building platform system based on a pre-trained large model includes modules such as: Agent (intelligent agent) development pipeline, business model management, application management service, data asset management service, model asset management service, data processing service, model management service, data development pipeline, and AI training and inference pipeline.

[0073] The Agent development pipeline includes modules for task decomposition / execution, feedback, Agent log management and monitoring, capability overview, Agent planning, and Agent organization.

[0074] The business model management module includes marketing, procurement, and supply chain management.

[0075] The application management service module includes organization, permissions, workflow, roles, Robotic Process Automation (RPA), engine, dynamic modeling, etc.

[0076] The data asset management service module includes topics, valuations, industry datasets, metadata, master data, etc.

[0077] Model asset management services include: AI atomic capabilities, AIGC large model ecosystem, etc.

[0078] Data processing services include functions and tools such as synchronization, integration, indexing, tagging, mining and development, and data asset management.

[0079] Model management services include: framework adaptation, data annotation, model training / evaluation, model development, service deployment and monitoring, model security, model encapsulation, distributed and high-performance systems, SFT, data version rollback, model version rollback, and more.

[0080] In addition, the aforementioned AI native application building platform system based on pre-trained large models may also include other auxiliary modules such as a platform security module.

[0081] Figure 6 The functions of the data asset management service, data processing service, and data development pipeline correspond to the functions of the data processing workflow module in the aforementioned embodiments.

[0082] Figure 6 The model asset management service, model management service, and AI training and inference pipeline in this document correspond to the functions of the model processing flow module in the aforementioned embodiments.

[0083] Figure 6 The Agent development pipeline module corresponds to the function of the multi-agent development pipeline module in the aforementioned embodiments.

[0084] Figure 6 The data asset management service and model asset management service in the document correspond to the functions of the aforementioned asset management module.

[0085] The AI ​​native application building platform system based on pre-trained large models provided in this invention has the following characteristics:

[0086] 1) Based on unified native capabilities, it can realize a loosely coupled, modular architecture for engineering pipelines, algorithms, data assets, business / process models and application systems, facilitating the assembly of applications and the deep integration of data, business and intelligence in vertical fields such as oil and gas and energy industries.

[0087] 2) It can support the standardized and modular integration of AI native applications.

[0088] 3) The platform can provide a closed loop of scene discovery, adaptive model selection, application assembly, continuous large model prompts, and AI native application construction.

[0089] 4) Through the separation of data and application and the architecture design that allows for both separation and integration, the AI ​​native application building platform system based on pre-trained large models can achieve better semantic conversion between AI native services and business and information systems, as well as more secure AIGC enterprise-level application building.

[0090] 5) Addressing the challenges of cross-modal and multi-scenario intelligent applications, this AI-native application building platform system, based on pre-trained large models, constructs a multi-agent cluster. This cluster possesses agile application building capabilities, including planning, recommendation, search, scheduling, prediction, classification, recognition, judgment, dialogue, intent decomposition, and application assembly. By building and managing this multi-agent cluster, the potential of LLM and traditional AI can be continuously unlocked, achieving adaptability to various application scenarios.

[0091] 6) In the AI ​​native application building platform system based on pre-trained large models, multiple agents in the agent cluster can be flexibly invoked, and the agent cluster can interoperate across organizations and multi-role systems.

[0092] 7) The AI ​​native application building platform system based on pre-trained large models supports each independent module to call LLM to publish and execute cross-agent group execution interaction instructions.

[0093] 8) The AI ​​native application building platform system based on pre-trained large models is designed with a self-supervised reflection mechanism, which provides the agent with the ability to reflect and optimize. Through adaptive learning and user-in-the-loop mode, it can perform human-computer interaction based on prompting engineering, continuously stimulate the potential of LLM and the good performance of agent clusters, and realize the continuous iterative operation of AI native applications.

[0094] Based on the same inventive concept, embodiments of the present invention also provide a method for constructing AI native applications based on pre-trained large models. This method constructs AI native applications based on the AI ​​native application construction platform system based on pre-trained large models described in the foregoing embodiments. Specific implementation methods can refer to the specific implementation methods of the aforementioned AI native application construction platform system based on pre-trained large models. Therefore, the implementation of this method can be found in the implementation of the foregoing method, and repeated details will not be elaborated further.

[0095] The AI ​​native application building platform system and method based on pre-trained large models provided in this invention can drive multi-type and multi-modal intelligent agent clusters. Combined with specific scenarios, it calls large language models to perform adaptive assembly. In complex scenarios in vertical domains, it provides a solution for building AI native applications, performing training or fine-tuning of intelligent agents and large language models, testing and verification, and concise deployment. It also achieves agile integration with existing enterprise business information systems. Compared with traditional agile development and waterfall development, the AI ​​native application building platform system and method based on pre-trained large models provided in this invention significantly improves upon traditional agile development and waterfall development methods in terms of application selection accuracy, application assembly efficiency, application iteration efficiency, large language model selection and scheduling, and the effectiveness of multi-agent management.

[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A platform system for building AI native applications based on pre-trained large models, characterized in that, include: The data processing workflow module provides a complete pipeline from data resource collection to data asset identification, supporting alignment, feature fusion, content fusion, and tag fusion of multiple modal data; it also supports toolchain capabilities for data asset processing. The model processing module is used to provide support for the development and training of artificial intelligence models; The asset management module provides various assets and asset management services, and offers on-demand access through multiple interface protocols. The multi-agent development pipeline module includes an agent cluster containing multiple agents, and realizes agent cluster management, task decomposition and execution, task feedback and operation monitoring, forming a closed-loop link that supports the full life cycle management of agent clusters. It also supports embedding intelligent agent clusters into enterprise business systems to enable interaction between multiple intelligent agents and business processes.

2. The system as described in claim 1, characterized in that, The asset management module specifically includes: a data asset module, an artificial intelligence model asset module, a business model module, and an application model management service module, and provides on-demand access to existing information systems through Restful, SOAP, websocket, and GraphQL protocols.

3. The system as described in claim 1, characterized in that, The data processing module is used to support data synchronization from heterogeneous data sources such as Mini / O, object storage, MySQL database, SQL Server, NiFi, and vector database; and to support alignment, feature fusion, content fusion, and tag fusion operations for structured or unstructured multimodal data such as speech, images, natural language, and tables.

4. The system as described in claim 1, characterized in that, The data processing module specifically includes: a visual label management component, a data mining and development module, which supports multiple data mining frameworks and is compatible with multiple data annotation and development languages, and is used to realize the fusion and extraction of multimodal data features.

5. The system as described in claim 1, characterized in that, The model processing module supports artificial intelligence model frameworks such as TensorFlow, PyTorch, PaddlePaddle, and Mindsphere. It provides data annotation and a visual interface to support training, evaluation, and debugging. After the model training is completed, it is published through its own inference platform.

6. The system as described in claim 1, characterized in that, The agents in the multi-agent development pipeline module are used to receive user requests issued through natural language, perform machine understanding on the user intent corresponding to the request, break down the task chain based on the understood user intent, form the corresponding AI native application, execute the work plan of the AI ​​native application and drive execution and application debugging. During the debugging process, it achieves link integration and semantic alignment with existing enterprise business systems. After the joint debugging of multiple agents is successful, an interface is generated, embedded into the existing enterprise business system, executes the work plan corresponding to the request, and returns the execution result to the user.

7. The system as described in claim 1, characterized in that, The joint debugging of multiple intelligent agents is achieved through self-supervised reflection and continuous optimization within the multiple intelligent agents.

8. The system as described in claim 1, characterized in that, Multiple agents can also be retrained and optimized through user feedback on execution results.

9. The system according to any one of claims 1-8, characterized in that, The system has the capability to construct AI-native applications that include planning, recommendation, search, scheduling, prediction, classification, recognition, judgment, dialogue, intent decomposition, adaptive application assembly, and multi-agent collaboration.

10. A method for building AI-native applications, characterized in that, This method constructs AI native applications using the AI ​​native application construction platform system based on pre-trained large models as described in any one of claims 1-9.

Images