Model application development method, system, computer device, readable storage medium and program product

Abstract

The application relates to a model application development method and system, computer equipment, a computer readable storage medium and a computer program product. The method realizes development of a model application through an application development support framework, specifically comprising the following steps: firstly, analyzing an application development request for a target application to obtain module selection information corresponding to the target application; then, based on the module selection information, configuring a generative language model on the basis of the application development support framework and configuring a data storage module; then, searching for a matched application framework; constructing a capability expansion tool module and an operation and maintenance support module; simultaneously, constructing a data preprocessing module and a retrieval enhancement module; and finally, constructing the target application on the basis of the application development support framework. The layered design of the model application framework and the model application architecture method based on the framework solve the gap between large models and actual application landing and the sedimentation of public technical components, and can effectively improve the research and development efficiency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a model application development method, system, computer device, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of computer technology and artificial intelligence, various types of large-scale generative language models have developed rapidly, with their main architecture based on Transform models. Generative large language models can specifically include general-purpose large language models, multimodal large models, and diffusion models, which can form a basic group of generative models. Furthermore, with the rapid development of the mobile internet, various application platforms have been subdivided into numerous sub-business scenarios, some of which involve the application of various forms of large language model algorithms based on artificial intelligence. In actual business operations, many scenarios require the use of large models to achieve intelligent business upgrades, such as supporting the rapid construction of custom workflows, combining knowledge bases, automatically retrieving logs, problem analysis, custom script plugins, and the implementation of personalized functions and applications.

[0003] However, due to the complexity of large language models, problems have arisen such as incomplete supporting tools, lack of support from application development frameworks, and difficulties in managing application development data, which have affected application development efficiency. Summary of the Invention

[0004] Therefore, it is necessary to provide a model application development method, system, computer equipment, computer-readable storage medium, and computer program product that can effectively improve the development efficiency of applications based on large-scale generative language models, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a model application development method applied to a server, the server including an application development support framework, the application development support framework including a storage and model layer, an application framework layer, a tool and operation and maintenance support layer, a data support layer, and a business application layer, the server including generative large model modules and information on various types of application frameworks, the method including:

[0006] Parse the application development request for the target application to obtain the module selection information corresponding to the target application;

[0007] Based on the module selection information, a generative language model is configured for the target application in the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data.

[0008] Based on the module selection information, an application framework that matches the data storage module and the generative language model is found at the application framework layer;

[0009] Based on the module selection information, capability extension tool modules and operation and maintenance support modules corresponding to the generative language model are constructed in the tool and operation and maintenance support layer.

[0010] Based on the module selection information, a data preprocessing module and a retrieval enhancement module for the generative language model are constructed in the data support and business application layer.

[0011] Based on the generative language model, the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module, the target application is constructed at the business application layer.

[0012] Secondly, this application also provides a model application development system, set on a server, the server including an application development support framework, the application development support framework including a storage and model layer, an application framework layer, a tool and operation and maintenance support layer, a data support layer and a business application layer, including:

[0013] The request parsing module is used to parse the application development request for the target application and obtain the module selection information corresponding to the target application.

[0014] The storage and model layer module is used to configure a generative language model for the target application in the storage and model layer based on the module selection information, and to configure a data storage module for the generative language model. The data storage module is used to store model dependency data.

[0015] An application framework layer module is used to search for an application framework that matches the data storage module and the generative language model in the application framework layer based on the module selection information.

[0016] The tools and operations support layer module is used to construct capability extension tool modules and operations support modules corresponding to the generative language model in the tools and operations support layer based on the module selection information.

[0017] The data support layer module is used to construct the data preprocessing module and retrieval enhancement module of the generative language model in the data support and business application layer based on the module selection information;

[0018] The business application layer module is used to construct the target application in the business application layer based on the generative language model, the data storage module, the application framework, the capability extension tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module.

[0019] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0020] Parse the application development request for the target application to obtain the module selection information corresponding to the target application;

[0021] Based on module selection information, a generative language model is configured for the target application at the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data.

[0022] Based on module selection information, the application framework layer is used to find application frameworks that match the data storage module and the generative language model.

[0023] Based on module selection information, capability extension tool modules and operation and maintenance support modules corresponding to generative language models are built in the tool and operation and maintenance support layer;

[0024] Based on module selection information, a data preprocessing module and a retrieval enhancement module for a generative language model are constructed at the data support and business application layers.

[0025] Based on generative language models, data storage modules, application frameworks, capability extension tools, operation and maintenance support modules, data preprocessing modules, and retrieval enhancement modules, the target application is built at the business application layer.

[0026] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0027] Parse the application development request for the target application to obtain the module selection information corresponding to the target application;

[0028] Based on module selection information, a generative language model is configured for the target application at the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data.

[0029] Based on module selection information, the application framework layer is used to find application frameworks that match the data storage module and the generative language model.

[0030] Based on module selection information, capability extension tool modules and operation and maintenance support modules corresponding to generative language models are built in the tool and operation and maintenance support layer;

[0031] Based on module selection information, a data preprocessing module and a retrieval enhancement module for a generative language model are constructed at the data support and business application layers.

[0032] Based on generative language models, data storage modules, application frameworks, capability extension tools, operation and maintenance support modules, data preprocessing modules, and retrieval enhancement modules, the target application is built at the business application layer.

[0033] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0034] Parse the application development request for the target application to obtain the module selection information corresponding to the target application;

[0035] Based on module selection information, a generative language model is configured for the target application at the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data.

[0036] Based on module selection information, the application framework layer is used to find application frameworks that match the data storage module and the generative language model.

[0037] Based on module selection information, capability extension tool modules and operation and maintenance support modules corresponding to generative language models are built in the tool and operation and maintenance support layer;

[0038] Based on module selection information, a data preprocessing module and a retrieval enhancement module for a generative language model are constructed at the data support and business application layers.

[0039] Based on generative language models, data storage modules, application frameworks, capability extension tools, operation and maintenance support modules, data preprocessing modules, and retrieval enhancement modules, the target application is built at the business application layer.

[0040] The aforementioned model application development methods, systems, computer equipment, computer-readable storage media, and computer program products, based on the application development support framework, can provide a unified data storage module and model deployment and management for the data and service dependencies required for model application development. First, the application development request for the target application is parsed to obtain the module selection information corresponding to the target application. The various module configurations for the target application by the developers are determined. Then, based on the module selection information, a generative language model is configured for the target application at the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data. Based on the module selection information, an application framework matching the data storage module and the generative language model is found at the application framework layer. Matching the application framework based on the determined data storage module and generative language model provides the architectural capabilities of a standard application framework. Directly based on the module selection information, capability extension tool modules and operation and maintenance support modules corresponding to the generative language model are built at the tool and operation and maintenance support layer. This provides some operation and maintenance support capabilities during the process of large model capability expansion and deployment service-oriented architecture, ensuring better operation and optimization of large model-related services. Based on module selection information, a data preprocessing module and a retrieval enhancement module for the generative language model are constructed at the data support and business application layer. This effectively provides the model's scalability in business scenarios, accelerates model iteration, and reduces development costs. Finally, based on the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module, the target application is built at the business application layer. This application provides a general architectural solution for model application architecture. The overall model application framework adopts a layered design to abstract, refine, and decompose the common demands and problems of various typical model applications: from the bottom up, it includes the storage and model layer, application framework layer, tool and operation and maintenance support layer, data support layer, and business application layer. The division of labor, core capabilities, and implementation methods of each layer are extended, thereby completing the management operations of application development on the basis of the generative large language model application development framework. The layered model application framework and the model application architecture method based on this framework provided by this application bridge the gap between large models and practical application implementation and the accumulation of common technical components, which can effectively improve R&D efficiency. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an application environment diagram of the model application development method in one embodiment;

[0043] Figure 2 This is a flowchart illustrating a model application development method in one embodiment;

[0044] Figure 3 This is a schematic diagram of a hierarchical support framework for rapid development of large-scale models in one embodiment.

[0045] Figure 4 This is a schematic diagram of the storage and model layer structure in one embodiment;

[0046] Figure 5 This is a schematic diagram of the structure of the tools and operation and maintenance support layer in one embodiment;

[0047] Figure 6 This is a schematic diagram of the data support layer in one embodiment;

[0048] Figure 7 This is a schematic diagram of the structure of a layered support framework for rapid development of large-scale models in another embodiment.

[0049] Figure 8 This is a schematic diagram of the structure of an application building system based on a generative large-scale language model for rapid application development and hierarchical support framework, as shown in one embodiment.

[0050] Figure 9 This is a structural block diagram of a model application development system in one embodiment;

[0051] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] The model application development method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A storage system can store the data that server 104 needs to process. The storage system can be integrated onto server 104 or located on a cloud or other network server. When application developers on terminal 102 want to develop applications based on a large language model, they can submit a corresponding application development request to server 104 through terminal 102, specifying the functional modules selected during model development in the request. Upon receiving the request from terminal 102, server 104 can construct the corresponding model application based on its model application development framework. The application development support framework includes a storage and model layer, an application framework layer, a tool and operation and maintenance support layer, a data support layer, and a business application layer, which can be used to support the development of model applications. Server 104 first parses the application development request for the target application to obtain the module selection information corresponding to the target application. Based on the module selection information, it configures a generative language model for the target application at the storage and model layer, and configures a data storage module for the generative language model. The data storage module is used to store model dependency data. Based on the module selection information, it searches for an application framework that matches the data storage module and the generative language model at the application framework layer. Based on the module selection information, it builds a capability extension tool module and an operation and maintenance support module corresponding to the generative language model at the tool and operation and maintenance support layer. Based on the module selection information, it builds a data preprocessing module and a retrieval enhancement module for the generative language model at the data support and business application layer. Based on the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module, it builds the target application at the business application layer. The target application can then be deployed to the corresponding application environment. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0054] In one exemplary embodiment, such as Figure 2 As shown, a model application development method is provided, which is applied to... Figure 1 Taking server 104 as an example, the server includes the application development support framework. For the application development support framework of large-scale applications, please refer to... Figure 3As shown, the framework includes a storage and model layer 302, an application framework layer 304, a tool and operation support layer 306, a data support layer 308, and a business application layer 310. The top layer of the framework is the business application layer 310, and the layers below it support the implementation of the layers above. This can be understood as the implementation of the functions of the upper layer being provided by the capabilities of the lower layer. For example, the application framework layer 304 is supported by the storage and model layer 302, and the data support layer is supported by the tool and operation support layer. In the specific development process, data flows from the upper application layer down to the bottom layer, and then from the bottom layer back to the top layer. When the server receives an application development request from a developer, the instruction data flows from the upper application layer down to the bottom layer, calling data at each layer to perform corresponding data processing operations. Finally, the processed result flows back from the bottom layer to the top layer, completing the entire instruction processing process. The upward arrows in the diagram represent this layered abstract relationship.

[0055] In this embodiment, the model application development method based on the application development support framework includes the following steps 201 to 209. Wherein:

[0056] Step 201: Server 104 parses the application development request for the target application and obtains the module selection information corresponding to the target application.

[0057] The application development request refers to a request submitted by an application developer to the server 104 via terminal 102, requesting the server 104 to develop a specified model application. The application development request can specify the application function modules to be used through module selection information. For example, it can specify the large language model to be used, and whether the model needs to be configured with embedding encoding. It can also specify modules that extend the model's functionality, such as modules that support the model's access to external data, and modules that support extended model functions (such as speech-to-text, text-to-image, etc.). Furthermore, the application development request can also specify information such as the application type and application scenario.

[0058] For example, this application is specifically applied to the efficient construction and generation of large-scale application models based on an application development support framework deployed on server 104. When staff on terminal 102 determine that a large-scale application model needs to be developed, they can first send an application development request to server 104 through 102, specifying the functional modules to be applied in the target application in the application development request. Server 104 then parses the application development request to obtain the module selection information corresponding to the target application, and thus completes the application development process based on the application development support framework. In a specific embodiment, when developers are developing based on the application development support framework, the first step is to determine the type of their large-scale application model, such as a role-based chatbot, document processing, chat session assistance, etc., and then sequentially select the support modules at each level of the framework. This order is from top to bottom from the data support layer 308. Developers need to first select the modules corresponding to the capabilities provided by the data support layer, then select the modules corresponding to the capabilities of the tool operation and maintenance support layer, and finally implement the corresponding modules of the tool application layer through the application layer framework. Finally, the large-scale application model layer is stored to provide underlying implementation support for the upper layers. Think of it as a Lego component library. Application developers select the modules and components they need from here to assemble, but the specific assembly method and its appearance are defined and selected by the developers based on their specific business scenarios. For example, a role-based chatbot requires data writing to the data support layer. If the extended tools only have basic capabilities and don't need to call third-party tools or complex tool combinations, they can be done through function modules. If they also need to call remote external capabilities, such as querying the weather or performing complex calculations, then remote function calls are required. On the other hand, if the model layer only needs a large language model, then the storage system is responsible for managing the various types of historical data subsequently generated by the model, and a key-value database is used to store the various types of historical data of the large language model.

[0059] In step 203, server 104 configures a generative language model for the target application in the storage and model layer based on the module selection information, and configures a data storage module for the generative language model. The data storage module is used to store model dependency data.

[0060] The storage and model layer includes various generative large-scale model modules, including LLM (Large Language Model), MLLM (Multilingual Large Language Model), and Diffusion Model. LLM refers to a computer model capable of processing and generating natural language; it represents a significant advancement in artificial intelligence and holds the promise of transforming the field through learned knowledge. LLM can predict the next word or sentence by learning the statistical patterns and semantic information of language data. Its capabilities increase as the input dataset and parameter space expand. It is used in various application areas such as robotics, machine learning, machine translation, speech recognition, and image processing, hence the name "multilingual large-scale language model." Diffusion models are generative models primarily used to generate high-quality images, audio, and other types of data. The core idea of ​​diffusion models is to gradually transform a simple distribution (usually Gaussian noise) into a complex data distribution, generating the final target data through a series of denoising steps. Furthermore, the storage and model layer also enables unified basic processing such as calling third-party large-scale model providers and quantization and acceleration during model deployment, reducing deployment costs. The data storage module refers to the functional module that provides data storage capabilities. In this application, the data storage module is specifically used to store various types of model-dependent data. Model-dependent data refers to the data required during model training and use, such as knowledge bases, dialogue history information, and knowledge graphs. The data storage module configured for model application development can include different data storage modules such as CFS file storage, KV (Key-Value Tree Storage), and HDFS document content storage, which can provide a unified storage access and permission management interface to the upper layer. The data storage module can interact with the generative language model, for example, by providing training data for the generative language model.

[0061] For example, after parsing the application development request, server 104 will build the target application of the large model from top to bottom based on the application development support framework. The first step is to complete the implementation of the model building foundation. That is, the configuration of the large model and data storage module is implemented in the storage and model layer. The storage and model layer is the foundation of the entire framework. It provides a unified data storage module and a model deployment and management layer for the data and service dependencies required by all large model applications. When an application development request is received, the generative language model required by the requesting application can be configured through the storage and model layer. After the model configuration is completed, the training and supervised fine-tuning of the model itself, after training alignment and debugging, are also managed in this layer. As for the data storage module, knowledge bases, dialogue history information, and knowledge graphs can be imported into different databases for storage. In one specific embodiment, a user wants to develop a question-answering robot for a news application. In this case, the user can use the model application development method of this application to develop a corresponding question-answering module for the news application and load it into the news application. At this time, the user can generate an application development request, specify the Large Language Model (LLM) as the base model in the request, and select the news database and news-related knowledge graph library as the database for the model application. Then, the Large Language Model is configured on the large model module, and the news database and news-related knowledge graph library are configured in the data storage module.

[0062] Step 205: Based on the module selection information, server 104 searches for an application framework that matches the data storage module and the generative language model at the application framework layer.

[0063] An application framework is a software tool or platform that provides the infrastructure and common functionality for developing specific types of applications. Application frameworks typically contain predefined classes, libraries, tools, and conventions to help developers build applications more efficiently, reduce repetitive code writing, and improve code maintainability and scalability. The application framework layer of this application mainly implements some standard frameworks in the application development process, such as interface call frameworks and interface naming frameworks.

[0064] For example, during application development, an application framework is also needed to provide the basic structure and general functions for the application development process, enabling developers to perform their work efficiently. Therefore, based on the module selection information, the application framework layer 304 can be used to configure the corresponding application framework for the target application to be developed. The configured framework includes standard frameworks, such as RPC (Remote Procedure Call) frameworks and various interface standard frameworks. The application framework can also be used to manage and maintain the MainIndex master index in the LangChain open-source components and vector database, ensuring application development efficiency.

[0065] Step 207: Based on the module selection information, server 104 constructs capability extension tool modules and operation and maintenance support modules corresponding to the generative language model in the tool and operation and maintenance support layer.

[0066] The capability extension tool module, part of the layered support framework for rapid development of large-scale model applications, is used to extend model functionality and consists of various capability extension tools. It primarily provides capability extensions for generative language models in the target application, ensuring better operation and optimization of large-scale model-related services. This mainly includes calls to third-party functions, primarily standardized capabilities such as interfaces to external search engines, speech-to-text services, and independent text-to-image services. The operations and maintenance support module, on the other hand, mainly provides operational support capabilities during the model deployment service-oriented process. It consists of several key parts: service capability development, observation, deployment, and evaluation. (Refer to...) Figure 3 As shown, the tool and operation support layer, which consists of the capability expansion tool module and the operation and maintenance support module, can provide the upper data support layer with the capability expansion tools and operation and maintenance support required for the development process.

[0067] For example, during application development, to achieve capability expansion and operational support for the generative language model, the tool and operational support layer 306 can be used to provide the capability expansion tools and operational support capabilities required for the model application. Therefore, based on module selection information, capability expansion tool modules and operational support modules corresponding to the generative language model can be constructed in the tool and operational support layer. These capability expansion tool modules enable the expansion of large model capabilities, including calls to third-party functions. These capability expansion tools primarily implement standardized capabilities such as interfaces to external search engines, speech-to-text service calls, and independent text-to-image service calls. Another part involves FunctionCall and Remote Function calls based on user intent recognition. Specific capability expansion tools can be selected according to the application's domain and application type. The operation and maintenance support module mainly provides support for the deployment and use of large models, including key parts such as the development, observation, deployment and evaluation of service capabilities. The development here mainly involves calling specific large models to implement relevant capabilities on the large model base model or the base model after fine-tuning, through prompt words and instructions. At the same time, it provides support for evaluating the actual performance of the large model, the number of input and output words (tokens), the way the output results are embedded in specific user interaction scenarios, and the actual performance of the deployed model service capabilities.

[0068] Step 209: Based on the module selection information, server 104 constructs a data preprocessing module and a retrieval enhancement module for the generative language model at the data support and business application layer.

[0069] The data preprocessing module primarily enhances the support for peripheral data dependent on the large model. For example, the short-term and long-term memory capabilities of intelligent dialogue agents can be managed offline to ensure data validity. The data preprocessing module consists of sub-modules for data acquisition, data processing, and data storage. The retrieval enhancement module retrieves relevant content for the large language model from the data managed by the data preprocessing module. It mainly consists of sub-modules for query term rewriting, intent recognition, relevance analysis, and filtering. (See reference...) Figure 3 As shown, the data support layer, composed of the data preprocessing module and the retrieval enhancement module, can call various functional modules of the tool and operation and maintenance support layer, and can also provide the data support required for the development process of the upper business application layer.

[0070] For example, in the application of large language models, adding new knowledge through fine-tuning is quite difficult, and adding new knowledge to the model also affects its existing knowledge. During fine-tuning, it's impossible to control which weights will be changed, and therefore impossible to know which capabilities will increase or decrease. Therefore, in many application scenarios that require efficiently expanding the knowledge capabilities of large models, importing external knowledge can ensure model performance. That is, through methods such as... Figure 3 The data support layer 308 shown is used to construct the data preprocessing module and retrieval enhancement module of the generative language model, thereby providing data support for the generative language model. The data preprocessing module mainly enhances the support of the external dependent data of the large model. The processing of the retrieval enhancement module is a recall and filtering process of the data processed by the data preprocessing module, including query rewriting (query terms here refer to various questions entered by users, as well as search requests), intent recognition, relevance judgment, retrieval and reordering, etc., which can efficiently retrieve relevant data content and ensure the smooth operation of the model. In one embodiment, this application is applied to the intelligent module of a chatbot. The chatbot is a continuous dialogue with a large language model. Since the large language model is stateless, it is necessary to accumulate and temporarily store the historical dialogue content. In subsequent dialogues, the relevant historical chat dialogues are also given to the chatbot. As the chat dialogue messages accumulate and become particularly numerous and messy over time, it is often necessary to recall some key chat information through some recall strategy to realize the processing of recorded information. At this time, the data preprocessing module can be used to organize this part of the data, and then the retrieval enhancement module can efficiently retrieve various types of information required for relevant questions.

[0071] Step 211: Server 104 builds the target application at the business application layer based on the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module.

[0072] For example, after server 104 has completed the construction of the modules required for the model application development process based on the application development support framework—that is, based on the modules configured or generated in the four-layer system of storage and model layer 302, application framework layer 304, tool and operation and maintenance support layer 306, and data support layer 308—the model application can be developed in the business application layer 310 according to the actual business scenario requirements. In the specific development process, after the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module are all prepared, the target application can be built in the business application layer based on the data interaction relationships between these modules. In specific embodiments, some very typical large model applications, such as question-answering robots based on document content support libraries, intelligent agent role dialogue robots, and dialogue robots with memory capabilities, are very easy to implement and iterate rapidly. The various basic functions and capabilities that require various assistance, service-orientation, and evaluation around the large model can all become part of the framework, effectively accelerating the implementation of model applications.

[0073] This application effectively enhances the knowledge and illusion-resistance capabilities of large models by introducing an additional data support layer. During the development and deployment of model application products, a significant amount of work is focused on data processing. Taking a question-answering model as an example, retrieval enhancement first retrieves relevant content from the database or external search engines, then provides an answer based on this. Alternatively, various internal knowledge documents (including .doc, .pdf, Markdown, tables, scanned documents, etc.) are preprocessed and stored in the database in a machine-readable and searchable format. Then, each time a question is asked, the application first matches and retrieves the most similar text blocks from the database based on the user's query terms. These document fragments are used as prompts, combined with the question, and fed to the large model to generate the answer. The entire process involves a variety of data processing tools, which are highly versatile and can be integrated into the framework's middleware platform for unified construction and reuse by multiple sub-businesses. Abstraction of other key modules, such as adding a dedicated long-term memory module to improve the memory capacity of large language models, enhances the development efficiency of AI model deployment applications and improves application effectiveness.

[0074] The aforementioned model application development method, based on the application development support framework, provides a unified data storage module and model deployment and management for the data and service dependencies required for model application development. First, it parses the application development request for the target application to obtain the corresponding module selection information. It determines the various module configurations for the target application by the developers. Then, based on the module selection information, it configures a generative language model for the target application at the storage and model layer, and configures a data storage module for the generative language model. This data storage module stores model dependency data. Based on the module selection information, it searches for an application framework that matches the data storage module and the generative language model at the application framework layer. Matching the application framework based on the determined data storage module and generative language model provides the architectural capabilities of a standard application framework. Directly based on the module selection information, it builds capability extension tool modules and operation and maintenance support modules corresponding to the generative language model at the tool and operation and maintenance support layer. This provides operational support capabilities for the large model capability expansion and deployment service process, ensuring better operation and optimization of large model-related services. Based on module selection information, a data preprocessing module and a retrieval enhancement module for the generative language model are constructed at the data support and business application layer. This effectively provides the model's scalability in business scenarios, accelerates model iteration, and reduces development costs. Finally, based on the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module, the target application is built at the business application layer. This application provides a general architectural solution for model application architecture. The overall model application framework adopts a layered design to abstract, refine, and decompose the common demands and problems of various typical model applications: from the bottom up, it includes the storage and model layer, application framework layer, tool and operation and maintenance support layer, data support layer, and business application layer. The division of labor, core capabilities, and implementation methods of each layer are extended, thereby completing the management operations of application development on the basis of the generative large language model application development framework. The layered model application framework and the model application architecture method based on this framework provided by this application bridge the gap between large models and practical application implementation and the accumulation of common technical components, which can effectively improve R&D efficiency.

[0075] In an exemplary embodiment, the generative language model includes a target large language model and a content encoding model. Step 201, configuring the generative language model for the target application at the storage and model layer based on module selection information, includes: based on the module selection information, searching for a base large language model and a content encoding model of the type specified in the application development request at the storage and model layer; pre-training the base large language model to obtain a pre-trained large model; performing supervised fine-tuning on the pre-trained large model to obtain the target large language model; and configuring the target large language model and content encoding model for the target application.

[0076] Among them, the basic large language models include LLM, MLLM, and diffusion models, while the content encoding model is a model that embeds content into the code, i.e., the embedding model. It is mainly used for preprocessing when performing vectorized retrieval of knowledge base or document content. The content encoding model can effectively improve the performance of the model.

[0077] For example, regarding the configuration process of a generative language model, this application focuses on providing an application development framework for the generative language model to realize application development. Therefore, the generative language model is the foundation of this solution, and model deployment is achieved by configuring the generative language model. Furthermore, model management can be further implemented. Therefore, during application development, based on module selection information, the basic large language model of the type specified in the application development request can be found in the storage and model layer. Simultaneously, the corresponding content encoding model needs to be found, providing the large language model with a common vectorized retrieval capability through the content encoding model. After model deployment, the large model training process needs to be completed. Large model training includes two stages: a pre-training stage and a supervised fine-tuning (SFT) training stage based on the specific task of the application model. After model selection, pre-training and supervised fine-tuning training are performed in the storage and model layer to obtain a target large language model that can be used for practical application tasks. The target large language model and content encoding model are then configured in the storage and model layer. In specific embodiments, unified basic processing such as calling third-party large model providers and quantization and acceleration during model deployment can also be implemented at the storage and model layers to reduce deployment costs.

[0078] In this embodiment, by configuring a basic large language model and a content encoding model, and completing segmented training of the large language model, a generative language model as the basis for the application can be effectively configured for the model application. At the same time, the content encoding model can provide data foundation services for the target large language model, ensuring the processing effect of the model.

[0079] In an exemplary embodiment, configuring the data storage module for the generative language model in step 201 includes: finding the model dependency data corresponding to the generative language model based on the application development request; performing sharding processing on the model dependency data according to the data type of the model dependency data to obtain sharded data, wherein the data types of the model dependency data include dependency knowledge base, historical data, knowledge graph and document data; and vectorizing and storing the sharded data in various databases to obtain the data storage module for the generative language model.

[0080] Among these, "dependent database" refers to the knowledge base relied upon in the business process, and "historical data" refers to historical data selected based on the domain of the target application to be developed. For example, when the target application is a question-answering robot in the medical field, historical data can be selected from dialogue data between doctors and patients in the medical field. A knowledge graph, also known as a knowledge domain visualization or knowledge domain mapping map, is a series of various graphs that display the development process and structural relationships of knowledge. It uses visualization technology to describe knowledge resources and their carriers, mining, analyzing, constructing, drawing, and displaying knowledge and their interrelationships. When applied to model applications, configuring a knowledge graph database can enhance the semantic understanding and context awareness capabilities of generative language models in the application, improve factual reasoning and logical consistency, provide rich background knowledge, and provide data support for multimodal fusion and personalization. Document data is used to support various document-type data for large models. Vectorized storage is a data storage and processing technology that represents data as high-dimensional vectors and uses these vectors for efficient computation and retrieval.

[0081] For example, while configuring a large language model, it is also necessary to provide model dependency data for the development of model applications. Model dependency data serves as the training foundation for the model, encompassing various languages, styles, and themes, enabling the model to understand and generate natural language content across a wide range of domains. This data can also be used for continuous model optimization. Furthermore, by introducing documentation related to the specific domains of the model dependency data, the model can better serve specific industries and provide personalized customization services for model applications. During model application development, the required data can be specified through application development requests, thereby retrieving the corresponding model dependency data for the generative language model. Specifically, model dependency data includes various types of data such as dependency knowledge bases, historical data, knowledge graphs, and document data. When configuring the data, it can be sharded according to data type, and the sharded data can be vectorized and stored in various databases, including different storage systems such as CFS file storage, KV (Key-Value Tree Storage), and HDFS document content storage, providing a unified storage access and permission management interface to the upper layer. This results in the data storage module for the generative language model.

[0082] In this embodiment, after the data is sharded according to the data type, the sharded data is saved in vectorized form to each database. Through various types of sharded data, strong data support can be effectively provided for the development of generative language models and model applications, ensuring the efficiency of model training, model use and application development processes.

[0083] In one specific embodiment, the structure of the storage and model layer can be referred to Figure 4As shown, it comprises two parts: a storage system 401 and a generative large model 402. The storage system includes a knowledge base, a historical database of dialogue content, various third-party storage solutions, a knowledge graph library, a message log library, and a vector library. The generative large model provides support for models such as LLM, MLLM, and diffusion models, while also offering services such as third-party model support, model management, model quantization, and embedded coding. During model application development, the corresponding generative language models and data storage modules can be configured through the storage and model layers, providing a storage system, model deployment, and a management layer for model application development.

[0084] In an exemplary embodiment, step 205 includes: based on module selection information, searching for various standard application frameworks corresponding to the data storage module and the generative language model at the application framework layer; based on application development requests, determining the application interface framework and the main index management framework corresponding to the data storage module and the generative language model; and using the standard application framework, the application interface framework, and the main index management framework as the application frameworks that match the data storage module and the generative language model.

[0085] For example, at the application framework layer, the main responsibility is to connect to various frameworks used in the application development process. Standard application frameworks refer to those universally applicable to all applications, such as RPC remote call frameworks, which can provide direct capability calls to the model as a service. The application interface framework refers to the standards and definition specifications for model applications, including API naming conventions, interface capability classification and descriptions, interface call examples, etc. It is mainly used for interface development during the application development process. The main index management framework is used for the management and maintenance of the MainIndex in the LangChain open-source component and vector database. LangChain is an open-source framework for building language model applications, providing a series of tools and components to simplify the development process. The vector database is an important component of LangChain, used for storing and retrieving high-dimensional vector data. In the vector database, the main index is a key part of management and maintenance, ensuring efficient data storage and retrieval.

[0086] In this embodiment, by introducing various model application development frameworks based on matching relationships at the application framework layer, the overall application development efficiency can be effectively improved.

[0087] In an exemplary embodiment, step 207, which involves constructing a capability extension tool module and an operation and maintenance support module corresponding to the generative language model in the tool and operation and maintenance support layer based on the module selection information, includes: obtaining tool function call information of the generative language model based on the module selection information; and constructing a capability extension tool module of the generative language model in the tool and operation and maintenance support layer based on the tool function call information.

[0088] For example, utility functions refer to functions used to extend the functionality of large models. This extension is primarily achieved through calls to third-party utility functions. These mainly include standardized capabilities such as interfaces to external search engines, speech-to-text services, and independent text-to-image services. Another part involves function calls (FunctionCall) and remote function calls (RFC) based on user intent recognition. Therefore, during development, when extending the capabilities of a generative language model, it is necessary to obtain relevant utility function call information. Based on this, the capability extension tools for the generative language model are configured, and corresponding capability extension tool modules are constructed. These modules are then used to extend the capabilities of the generative language model during application development, improving the efficiency of model application development. In a specific embodiment, for function calls based on user intent recognition and remote function calls, the implementation requires using the Function Schema to define and describe the task's external capability calls. Specifically, this involves defining data standards or classification task schema definitions (or directly understanding them as the parameters and descriptions of the Function), so that the generative language model can predict the specific values ​​to be assigned to the Function parameters based on the provided Function.

[0089] Furthermore, the capability extension tool module for constructing generative language models based on tool function call information includes: determining the mutual call relationships and function orchestration / combination relationships between tool functions based on tool function call information; and constructing the capability extension tool module for generative language models based on the mutual call relationships and function orchestration / combination relationships between tool functions. Building upon tool function calls, more complex functions can be implemented through mutual calls and orchestration / combination between tool functions, and these functions will also be added to the capability extension tool module. For example, in one embodiment, if a user and a robot want to check the recent weather, after the generative language model understands the meaning of the text, it can call a remote weather information service interface, then pass the obtained data as parameters to a service that displays the weather using formatted icons, and finally present the final weather display result. This may involve multiple function calls and their mutual orchestration / combination, all of which can be supported by the capability extension tool module.

[0090] In this embodiment, a generative capability extension tool module is constructed based on the function call information related to the generative language model. This module expands the model's capabilities during the application development process, effectively improving development efficiency. Furthermore, the module supports function calls and their arrangement and combination, further extending the processing capabilities of the generative language model and ensuring the practical effectiveness of model applications.

[0091] In an exemplary embodiment, step 207, which involves constructing an operation and maintenance support module for the generative language model in the tool and operation and maintenance support layer based on module selection information, includes: determining operation and maintenance support information for the deployment and use process of the generative language model based on module selection information; constructing a development sub-module, an observation sub-module, a deployment sub-module, and an evaluation sub-module for model operation and maintenance in the tool and operation and maintenance support layer based on the operation and maintenance support information; and constructing the operation and maintenance support module based on the development sub-module, the observation sub-module, the deployment sub-module, and the evaluation sub-module.

[0092] For example, the Operations (Ops) module primarily provides support during the deployment and use of large models, ensuring better operation and optimization of related services. The operational support process for generative language models during deployment and practical application mainly includes several key parts: service capability development, observation, deployment, and evaluation. Therefore, when implementing the Operations support module, corresponding development, observation, deployment, and evaluation sub-modules can be built for each key part of the operational support process to implement various sub-operational functions during model deployment and use. Through the functional support of these sub-modules, overall operational support is achieved. For example, the development sub-module mainly uses prompt commands to call specific large models on the large model base model or a fine-tuned base model to implement related capabilities. It also provides evaluation support for the actual performance of the large model, the number of input and output tokens, the embedding method of output results and specific user interaction scenarios, and the actual performance of the deployed model's service capabilities. In a specific embodiment, the model operation and maintenance submodule may also include an application embedding submodule to enable the embedding of applications such as robot plugins and web page plugins.

[0093] In this embodiment, by using operation and maintenance support information, development sub-modules, observation sub-modules, deployment sub-modules, and evaluation sub-modules for model operation and maintenance are created. This enables the provision of various related support during the deployment and use of large models in the process of model application development, ensuring that model-related services can be better operated and optimized during the model application development process, thereby improving development efficiency.

[0094] In one embodiment, the overall structure of the tools and operations support layer can be referred to Figure 5As shown, the system comprises two parts: the extension tools section 501 and the operation and maintenance support section 502. The extension tools section 501 includes three types of extension tools: Functions, FunctionCall, and Remote Function Call. Functions enable extended functions such as search engine API calls, speech-to-text conversion, knowledge graph extraction, and text-to-graph generation. FunctionCall supports service-oriented frameworks and models, while Remote Function Call supports single-function and multi-function orchestration. The operation and maintenance support section 502 includes development, observation, deployment, evaluation, and application embedding submodules. The development submodule provides visual development, prompt word engineering, and development environment (Playground) support. The observation submodule provides various observation-related operation and maintenance functions, including observation of token consumption, model cost, trajectory, alarm / notification / data reporting, and time consumption. The application embedding module enables the embedding and management of functional plugins, including robot plugins, web plugins, and chat components. The deployment submodule is responsible for the dynamic scaling and fixed quota operation and maintenance management of the model deployment process. As for the evaluation submodule, it can use recall / precision or realism as evaluation indicators, construct datasets through business uploads or automatic collection, implement evaluation tools in an automated or manual manner, and be responsible for collecting user feedback to complete the evaluation-related management of the model operation and maintenance process.

[0095] In an exemplary embodiment, step 209, which involves constructing a generative language model configuration data preprocessing module at the data support and business application layer based on module selection information, includes: obtaining historical content data of the business scenario corresponding to the target application based on module selection information; determining the offline recall strategy corresponding to the historical content data based on the data type of the historical content data; and generating a data preprocessing module at the data support and business application layer to preprocess the historical content data based on the offline recall strategy.

[0096] For example, this application can enhance the support for peripheral dependency data of a large model through offline data preprocessing. In this case, the data preprocessing module can be configured to manage peripheral dependency data support for the current model application. Unlike the basic model dependency data managed by the data storage module, adding new knowledge through fine-tuning during the application of a large model is quite difficult, and the drawback is that adding new capabilities through fine-tuning will affect its existing knowledge. This is because during fine-tuning, it is impossible to control which weights will be changed, and therefore impossible to know which capabilities will be increased or decreased. However, in many application scenarios that require efficient expansion of the knowledge capabilities of large models, such as knowledge question-answering assistants, offline data preprocessing can provide data support for relevant applications in these specific vertical fields. Furthermore, through offline data retrieval strategies, key data that needs to be processed offline can be collected. For example, in a chatbot scenario, it can manifest as a continuous dialogue with a large language model. Since the large language model is stateless, it is necessary to accumulate and temporarily store historical dialogue content. In subsequent dialogues, the relevant historical chat content is also given to the chatbot. As the chat messages accumulate over time and become particularly numerous and disorganized, it is often necessary to recall some key chat information through a certain recall strategy, thereby realizing the processing of recorded information. That is, by specifying corresponding offline recall strategies for different historical content data, the corresponding data is recalled and preprocessed through the offline recall strategy for later use.

[0097] In this embodiment, a data preprocessing module is introduced to recall and preprocess historical data. This module stores various internal knowledge related to specific vertical fields in the database in a machine-readable and searchable format, thereby enhancing the support capability of the large model's peripheral dependent data, ensuring the practical effect of the model application, and improving application development efficiency.

[0098] In an exemplary embodiment, step 209, which involves constructing a retrieval enhancement module for a generative language model at the data support and business application layer based on module selection information, includes: constructing various levels of retrieval sub-modules for a generative language model at the data support and business application layer based on module selection information; and constructing a retrieval enhancement module based on the various levels of retrieval sub-modules.

[0099] The various retrieval sub-modules include query term rewriting, intent recognition, relevance analysis, filtering, retrieval processing, re-ranking, and retrieval fusion modules, each corresponding to a different stage of retrieval enhancement processing.

[0100] For example, the retrieval enhancement module primarily works in conjunction with the data preprocessing module. Through retrieval enhancement, each time an application asks a question, it first matches the most similar text blocks in the database based on the user's query. These document fragments are then used as prompts, combined with the question, and fed into the large model to generate an answer. The retrieval enhancement process utilizes a variety of data processing tools, which are highly versatile and can be centrally managed within the framework platform for unified construction and reuse across multiple sub-businesses. In specific application scenarios, these data processing tools can be invoked to construct various levels of retrieval sub-modules within the generative language model, thereby achieving corresponding retrieval enhancement functions through the interconnection of these sub-modules. This provides additional context for the large language model to generate better and more specific responses. This typically includes query rewriting (where "query" here broadly refers to various user-input questions and search requests), intent recognition, relevance assessment, retrieval, and re-ranking. Due to the involvement of different types of data, the retrieval results also support multi-path parallel processing, followed by merging and fusion. These steps can be implemented through the aforementioned retrieval sub-modules at each level, ensuring the effectiveness of retrieval enhancement.

[0101] In this embodiment, by constructing retrieval sub-modules at various levels of the generative language model, the processing of each stage of the retrieval enhancement process is realized, thereby enhancing the support capability of the large model's peripheral dependent data, ensuring the model's application effect, and improving application development efficiency.

[0102] In one embodiment, the structure of the data support layer can be referred to Figure 6As shown, the system comprises two parts: a data writing section (601) and a recall and filtering section (602), enabling data preprocessing and enhanced data retrieval. The data writing section includes three sub-modules: data acquisition, data processing, and data storage. The data acquisition sub-module acquires data such as PDFs, PPTs, docs, tables, and serialized data, and is responsible for connecting to the file system / database / stream. The data processing module processes different data types such as paragraphs, tables, chapters, formatting, images, and audio, and is also responsible for data slicing and custom data cleaning. The data storage module can store the processed data using methods such as embedded encoding, inverted indexes, and graph databases. The recall and filtering process includes sub-modules such as query term rewriting, intent recognition, relevance analysis, filtering, retrieval processing, re-ranking, and retrieval fusion. Query term rewriting can be implemented using a general or custom solution. The intent recognition and relevance processing processes can also use general or custom models. Filtering can be implemented using pre-filtering, post-filtering, or security filtering methods. For retrieval processing, multi-source data, metadata, and relevance analysis are employed. The re-ranking process utilizes a sorter, a general model, or a custom model. Finally, retrieval fusion is achieved through retrieval link orchestration: re-ranking vectors, text, and images.

[0103] In an exemplary embodiment, step 211 includes: obtaining the data interaction relationship between the generative language model and data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module; based on the data interaction relationship, connecting the generative language model and data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module, and constructing the target application at the business application layer.

[0104] For example, after configuring and organizing the functional modules of the model application development process through the application development support framework, and configuring the generative language model, data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module for the target application, the data interaction relationships between these modules—that is, the data interaction relationships between the generative language model and the data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module—can be used to connect the generative language model with the data storage module, application framework, capability extension tool module, operation and maintenance support module, data preprocessing module, and retrieval enhancement module, thereby constructing the target application at the business application layer. Figure 3As shown, the model is constructed specifically in the business application layer 310, supported by four layers of systems: storage and model layer 302, application framework layer 304, tool and operation and maintenance support layer 306, and data support layer 308. The model construction process may also require development based on the application type and scenario of the target application. For example, question-answering robots based on document content support libraries, intelligent agent role-based dialogue robots, and dialogue robots with memory capabilities can all be considered as target applications. Application development and optimization also need to be combined with specific application scenarios. That is, after determining the application type and scenario of the target application, the target application corresponding to the application type is generated, and the target application is deployed to the application scenario to ensure efficient application deployment.

[0105] In this embodiment, based on the application development support framework, various functional modules are constructed. These modules are then connected through data interaction and association using a generative language model to build the target application. Furthermore, the architecture and development management of the model application can be implemented based on application type and application scenario. Simultaneously, it can fully utilize the diverse knowledge and rich logical reasoning capabilities contained in the large language model to efficiently process knowledge-related content and application scenarios such as knowledge question answering and knowledge reasoning, thereby enhancing the scalability of AI model applications. Specialized components such as search, memory, and operational support provide support to accelerate application deployment through rapid testing and iteration, reducing repetitive development work.

[0106] This application also provides an application scenario, which is illustrated by applying the above-mentioned model application development method to this application scenario. The model application development method specifically includes:

[0107] When users want to develop an intelligent question-answering robot for a video application, they hope to use a generative language model as a foundation to ensure the accuracy of the model's question-answering. In this case, users can use the model application development method described in this application to manage the application development process, ensuring the efficiency of the intelligent question-answering robot development and the final question-answering results. This application specifically utilizes, for example... Figure 7 The generative large model application rapid development and support system shown adopts a layered design for the overall AI model application framework. It abstracts, refines and decomposes the common demands and problems of various typical AI model applications. It includes multiple layers from the bottom up, such as storage and model layer, model application framework layer, tool and operation and maintenance support layer, data support layer and business application layer.

[0108] During application development, developers first need to configure various functional modules for the intelligent question-answering robot, which is the development target. The system then completes the configuration of application modules and application generation processing based on the application development support framework submitted by the developers. The system first configures the corresponding generative language model and data storage module at the storage and model layer. At this time, an LLM model can be selected, and through the data storage module, knowledge bases, knowledge graphs, message records, etc. related to video content and video applications are set as model dependency data for the intelligent question-answering robot.

[0109] Then, at the application framework layer, corresponding application frameworks are matched for the development of intelligent question-answering robots, including various standard application frameworks, such as RPC and AI application interface frameworks and the management framework of the MainIndex in LangChain open-source components & vector database.

[0110] In the tools and operations support layer, which serves as a crucial intermediary layer connecting the upper and lower layers of the framework, the main function is to provide operational support capabilities for the expansion and service-oriented deployment of large models. This ensures better operation and optimization of services related to large models. Specifically, it implements capability expansion management and operations support for intelligent question-answering robots. For intelligent question-answering robots, it can obtain tool function call information for generative language models; then, based on this information, it constructs capability expansion tool modules for generative language models. Furthermore, it builds development, observation, deployment, and evaluation sub-modules for model operations; and then, based on these sub-modules, it constructs operations support modules.

[0111] The data support layer mainly consists of two parts: offline data preprocessing and retrieval enhancement that combines data with the large model. The former primarily enhances the support of the data that the large model depends on. Then, through the retrieval enhancement module, during the dialogue with the intelligent agent, relevant parts are retrieved from historical dialogue information as context for the final response. This provides additional context for the large language model (LLM) to generate better and more specific responses. For the development of intelligent question-answering robots, historical content data related to the question-answering scenario of video content is acquired; based on the data type of the historical content data, the corresponding offline recall strategy is determined; and then, a data preprocessing module is generated based on the offline recall strategy to preprocess the historical content data. Simultaneously, a retrieval enhancement module is generated based on modules such as query term rewriting, intent recognition, relevance analysis, filtering, retrieval processing, re-ranking, and retrieval fusion.

[0112] Finally, at the business application layer, supported by the four layers of the system described above, the development and processing of the question-answering robot can be realized, and the model can be continuously and rapidly iterated. Furthermore, the various basic functions and capabilities that need to be supported by the large model, such as various auxiliary functions, service-oriented functions, and evaluation functions, can all become part of the framework, improving the development efficiency of model applications, and finally completing the development and management of the intelligent question-answering robot.

[0113] The model application development method implemented through the above-mentioned generative large model application enhancement R&D framework can fully utilize the various knowledge and rich logical reasoning capabilities contained in large language models to efficiently process knowledge-related content and application scenarios such as knowledge question answering and knowledge reasoning, improve the scalability of AI model applications, and accelerate the application deployment through specialized components such as search, memory, and operation and maintenance support, enabling rapid trial and iteration and reducing repetitive development work.

[0114] At the same time, it can effectively provide and accumulate public technical components in the process of applying large AI models, such as providing unified interface retrieval, storage and query capabilities, model result evaluation capabilities, data preprocessing capabilities, etc., effectively accelerating the R&D iteration speed of upper-layer applications based on generative large language models, reducing the cost of R&D in multiple scenarios, and lowering the technical threshold for the implementation of AI applications.

[0115] Finally, it can also effectively improve the support efficiency of project implementation and deployment. By building a unified basic support platform, unified service interfaces and service deployment, monitoring, log collection and analysis, etc., and adding a dedicated long-term memory module to extend the capabilities of large language models and provide good linkage support for the adaptability and application effects of various business development.

[0116] In one embodiment, this application also provides a hierarchical support system for rapid application development based on generative large-scale language models, outlining its relationship and role with other modules in a typical instant messaging social network system. In this scenario, a structural diagram of the application construction system based on the hierarchical support framework for rapid application development using generative large-scale language models can be found [here]. Figure 8 As shown, the functions of each service module included are as follows:

[0117] I. Content Production and Consumption

[0118] (1) PGC or UGC (user generated content), MCN content producers provide text or video content through mobile terminals or backend API systems. These are the main sources of content for recommended distribution.

[0119] (2) Upload text and image content through communication with upstream and downstream content interface services. The source of text and image content is usually a lightweight publishing end and editing content entry. Video content publishing is usually a shooting end. During the shooting process, local video content can choose to match music, filter templates and video beautification functions, etc.

[0120] (3) As a consumer, I communicate with the upstream and downstream content interface server, push the recommendation to obtain the index information of the accessed content, and then communicate with the content storage server to obtain the corresponding content, including the recommended content and the content of the topic subscription. The content storage server stores the content entities such as video source files and image source files, while the content meta information such as title, author, cover image, category, tag information, etc. are stored in the content database.

[0121] (4) The client includes various typical applications based on the AI ​​big model, such as intelligent agent chatbots, chat group document summaries, chat group message summaries and assistants, knowledge-enhanced intelligent agents, vertical knowledge base question answering robots, chat channel auxiliary assistants, etc., which communicate directly with the services deployed by the big model on the client and back end.

[0122] II. Uplink and Downlink Content Interface Server

[0123] (1) Communicate directly with the content production end. The content submitted from the front end is usually the title, publisher, summary, cover image, and publication time of the content. The file is stored in the content database. The large model application of the end also completes related communication with the server through the upstream and downstream interface server.

[0124] (2) The submitted content is synchronized to the scheduling center server for subsequent content processing and transfer.

[0125] III. Content Database

[0126] (1) The core database of content, where all the metadata of the content published by producers is stored. The focus is on the metadata of the content itself, such as file size, cover image link, bitrate, file format, title, publication time, author, video file size, video format, whether it is original or first release, as well as the classification of content during the manual review process (including first, second and third level classification and tag information, for example, an article explaining A brand mobile phone, the first level classification is technology, the second level classification is smartphone, the third level classification is domestic mobile phone, and the tag information is A brand, X model, shooting artifact);

[0127] (2) The content processing of the dispatch center mainly includes machine processing and manual review. The core of machine processing is various quality judgments, such as low-quality filtering, content tags such as classification, tag information, and content deduplication. Their results will be written into the content database. Completely duplicate content will not be given to manual secondary processing.

[0128] (3) Subsequent applications, such as content summarization and generalization, can use the content in the content database as the basis for contextual reasoning in large model processing.

[0129] IV. Dispatch Center & Manual Review System

[0130] (1) Responsible for the entire scheduling process of content flow, receiving the content to be stored in the database through the upstream and downstream content interface servers, and then obtaining the content's metadata from the content database;

[0131] (2) The manual review system is the carrier of human service capabilities. It is mainly used to review and filter sensitive, pornographic, and legally prohibited content that machines cannot determine. At the same time, it also performs video content labeling and secondary confirmation.

[0132] (3) During the manual review process, the results output by the AI ​​large model application, such as the results of dialogue responses, the results of knowledge assistant answers, etc., are evaluated, scored and marked. The relevant data can be used as sample data for the corresponding application iteration and update.

[0133] V. Content Storage Service

[0134] (1) Usually, it is a group of storage servers that are widely distributed and located far away from users on the C side. There are usually CDN acceleration servers on the periphery for distributed caching acceleration. The video content uploaded by the content producer is saved through the uplink and downlink content interface servers.

[0135] VI. Business Sample Database

[0136] (1) Read business-related sample data from the content metadata service marked by manual review, such as high-quality Q&A content generated by Q&A assistant replies, and summaries and summaries of document or file content;

[0137] (2) Provide specific business scenario data for the application of AI large model in the corresponding business scenario for model fine-tuning and optimization.

[0138] VII. Large-scale language model application development support platform

[0139] (1) According to the above model application development method and application rapid development layered support framework description, through the layered design, from the bottom up, there are 5 major layers including storage and model layer, AI application framework layer, tool and operation and maintenance support layer, data support layer and business application layer. Then, the division of labor and the core capabilities and implementation methods of each layer are expanded.

[0140] (2) AI large-scale model applications in different business scenarios are developed and evaluated based on this unified framework, effectively improving and enhancing R&D efficiency. VIII. Reporting and Analysis Interface Services

[0141] (1) Communicate with the end to receive various feedback during the process of consuming and distributing the reported messages, such as feedback on the results of generating the AI ​​large model, which is used to evaluate the effectiveness of the AI ​​model in actual business and collect data for subsequent model improvement.

[0142] IX. Deduplication Service

[0143] (1) The main method is to deduplicat the content by using the fingerprint of the content. For video content, the video embedding encoding is usually used, and for text and image content, the content simhash is usually used.

[0144] 10. Data processing groups specific to business scenarios

[0145] (1) Here, the generative big model has different applications in different business scenarios, such as knowledge question answering assistant, content text summary and summary, intelligent agent role dialogue memory retrieval, etc. The data style and format in each scenario are processed accordingly, including data cleaning, formatting, sensitive word filtering, etc. Different data preprocessing to obtain specific data for specific business scenarios, and then these data are incorporated into the application rapid development layered support platform.

[0146] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0147] Based on the same inventive concept, this application also provides a model application development system for implementing the model application development method described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations in one or more model application development system embodiments provided below can be found in the limitations of the model application development method described above, and will not be repeated here.

[0148] In one exemplary embodiment, such as Figure 9 As shown, a model application development system is provided, hosted on a server. The server includes an application development support framework, which comprises a storage and model layer, an application framework layer, a tool and operation and maintenance support layer, a data support layer, and a business application layer, including:

[0149] The request parsing module 902 is used to parse the application development request for the target application and obtain the module selection information corresponding to the target application.

[0150] The storage and model layer module 904 is used to configure a generative language model for the target application in the storage and model layer based on module selection information, and to configure a data storage module for the generative language model. The data storage module is used to store model dependency data.

[0151] Application framework layer module 906 is used to find application frameworks that match the data storage module and generative language model in the application framework layer based on module selection information.

[0152] The Tools and Operations Support Layer Module 908 is used to build capability extension tool modules and operations support modules corresponding to the generative language model in the Tools and Operations Support Layer based on module selection information.

[0153] The data support layer module 910 is used to construct a data preprocessing module and a retrieval enhancement module for a generative language model in the data support and business application layers based on module selection information.

[0154] The business application layer module 912 is used to build target applications in the business application layer based on generative language models, data storage modules, application frameworks, capability extension tool modules, operation and maintenance support modules, data preprocessing modules, and retrieval enhancement modules.

[0155] In one embodiment, the generative language model includes a target large language model and a content encoding model. The storage and model layer module 904 is specifically used for: based on module selection information, searching the storage and model layer for the basic large language model and content encoding model of the type specified in the application development request; pre-training the basic large language model to obtain a pre-trained large model; supervising and fine-tuning the pre-trained large model to obtain the target large language model; and configuring the target large language model and content encoding model for the target application.

[0156] In one embodiment, the storage and model layer module 904 is further configured to: locate the model dependency data corresponding to the generative language model based on the application development request; perform sharding processing on the model dependency data according to the data type of the model dependency data to obtain sharded data, wherein the data type of the model dependency data includes dependency knowledge base, historical data, knowledge graph and document data; and vectorize and store the sharded data in various databases to obtain the data storage module of the generative language model.

[0157] In one embodiment, the application framework layer module 906 is specifically used to: search for various standard application frameworks corresponding to the data storage module and the generative language model in the application framework layer based on module selection information; determine the application interface framework and the main index management framework corresponding to the data storage module and the generative language model based on the application development request; and use the standard application framework, the application interface framework, and the main index management framework as the application frameworks for matching the data storage module and the generative language model.

[0158] In one embodiment, the tool and maintenance support layer module 908 is specifically used to: obtain tool function call information of the generative language model based on module selection information; and construct a capability extension tool module for the generative language model based on the tool function call information in the tool and maintenance support layer.

[0159] In one embodiment, the tool and operation support layer module 908 is further configured to: determine the mutual calling relationship and function orchestration and combination relationship between tool functions based on tool function call information; and, based on the mutual calling relationship and function orchestration and combination relationship between tool functions, construct a capability extension tool module for generative language models in the tool and operation support layer.

[0160] In one embodiment, the tool and operation and maintenance support layer module 908 is further configured to: determine operation and maintenance support information for the deployment and use of the generative language model based on module selection information; construct a development sub-module, an observation sub-module, a deployment sub-module, and an evaluation sub-module for model operation and maintenance in the tool and operation and maintenance support layer according to the operation and maintenance support information; and construct an operation and maintenance support module based on the development sub-module, the observation sub-module, the deployment sub-module, and the evaluation sub-module.

[0161] In one embodiment, the data support layer module 910 is specifically used to: obtain historical content data of the business scenario corresponding to the target application based on module selection information; determine the offline recall strategy corresponding to the historical content data according to the data type of the historical content data; and generate a data preprocessing module for data preprocessing of historical content data in the data support and business application layer based on the offline recall strategy.

[0162] In one embodiment, the data support layer module 910 is further configured to: construct various levels of retrieval sub-modules of the generative language model in the data support and business application layer based on module selection information; and construct a retrieval enhancement module based on the various levels of retrieval sub-modules.

[0163] In one embodiment, the business application layer module 912 is specifically used to: obtain the data interaction relationship between the generative language model and data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module; and based on the data interaction relationship, connect the generative language model and data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module to build the target application in the business application layer.

[0164] Each module in the above-mentioned model application development system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0165] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data related to model application development. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a model application development method.

[0166] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0167] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0168] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0169] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0172] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0173] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for developing model applications, characterized in that, Applied to a server, the server including an application development support framework, the application development support framework including a storage and model layer, an application framework layer, a tool and operation and maintenance support layer, a data support layer, and a business application layer, the method includes: Parse the application development request for the target application to obtain the module selection information corresponding to the target application; Based on the module selection information, a generative language model is configured for the target application in the storage and model layer, and a data storage module is configured for the generative language model. The data storage module is used to store model dependency data. Based on the module selection information, an application framework that matches the data storage module and the generative language model is found in the application framework layer; Based on the module selection information, capability extension tool modules and operation and maintenance support modules corresponding to the generative language model are constructed in the tool and operation and maintenance support layer. Based on the module selection information, a data preprocessing module and a retrieval enhancement module for the generative language model are constructed in the data support and business application layer. Based on the generative language model, the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module, the target application is constructed at the business application layer.

2. The method according to claim 1, characterized in that, The generative language model includes a target large language model and a content encoding model; The step of configuring a generative language model for the target application in the storage and model layer based on the module selection information includes: Based on the module selection information, the basic large language model and content encoding model of the application development request type are searched in the storage and model layer; The basic large language model is pre-trained to obtain a pre-trained large model; The pre-trained large model is subjected to supervised fine-tuning to obtain the target large language model, and the target large language model and the content encoding model are configured for the target application.

3. The method according to claim 1, characterized in that, The configuration of the data storage module for the generative language model includes: Based on the application development request, find the model dependency data corresponding to the generative language model; Based on the data type of the model dependency data, the model dependency data is segmented to obtain segmented data. The data types of the model dependency data include dependency knowledge base, historical data, knowledge graph and document data. The fragmented data is vectorized and stored in various databases to obtain the data storage module of the generative language model.

4. The method according to claim 1, characterized in that, The step of finding an application framework that matches the data storage module and the generative language model at the application framework layer based on the module selection information includes: Based on the module selection information, the application framework layer searches for various standard application frameworks corresponding to the data storage module and the generative language model. Based on the application development request, determine the application interface framework and main index management framework corresponding to the data storage module and the generative language model; The standard application framework, the application interface framework, and the main index management framework are used as the application framework for matching the data storage module and the generative language model.

5. The method according to claim 1, characterized in that, The step of constructing capability extension tool modules and operation and maintenance support modules corresponding to the generative language model in the tool and operation and maintenance support layer based on the module selection information includes: Based on the module selection information, obtain the tool function call information of the generative language model; The tool and maintenance support layer includes a capability extension tool module that constructs the generative language model based on the tool function call information.

6. The method according to claim 5, characterized in that, The capability extension tool module for constructing the generative language model based on tool function call information in the tool and operation support layer includes: Based on the utility function call information, determine the mutual call relationships and function arrangement and combination relationships between utility functions; Based on the mutual calling relationships and function orchestration and combination relationships among the tool functions, a capability extension tool module for building the generative language model is constructed in the tool and operation and maintenance support layer.

7. The method according to claim 1, characterized in that, Based on the module selection information, the operation and maintenance support module for constructing the generative language model in the tool and operation and maintenance support layer includes: Based on the module selection information, the operation and maintenance support information for the deployment and use of the generative language model is determined; Based on the operation and maintenance support information, a development sub-module, an observation sub-module, a deployment sub-module, and an evaluation sub-module for model operation and maintenance are constructed in the tool and operation and maintenance support layer. An operation and maintenance support module is constructed based on the development submodule, the observation submodule, the deployment submodule, and the evaluation submodule.

8. The method according to claim 1, characterized in that, Based on the module selection information, the generative language model configuration data preprocessing module constructed in the data support and business application layer includes: Based on the module selection information, historical content data of the business scenario corresponding to the target application is obtained; Based on the data type of the historical content data, determine the offline recall strategy corresponding to the historical content data; Based on the offline recall strategy, a data preprocessing module is generated at the data support and business application layer to preprocess the historical content data.

9. The method according to claim 8, characterized in that, Based on the module selection information, the retrieval enhancement module for constructing the generative language model in the data support and business application layer includes: Based on the module selection information, the generative language model is constructed at various levels of retrieval sub-modules in the data support and business application layer; Based on the aforementioned retrieval submodules at each level, a retrieval enhancement module is constructed.

10. The method according to any one of claims 1 to 9, characterized in that, The construction of the target application at the business application layer based on the generative language model, the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module includes: Obtain the data interaction relationships between the generative language model and the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module; Based on the data interaction relationship, the generative language model is connected to the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module to construct the target application at the business application layer.

11. A model application development system, characterized in that, The system is hosted on a server, which includes an application development support framework. This framework comprises a storage and model layer, an application framework layer, a tool and operations support layer, a data support layer, and a business application layer. The request parsing module is used to parse the application development request for the target application and obtain the module selection information corresponding to the target application. The storage and model layer module is used to configure a generative language model for the target application in the storage and model layer based on the module selection information, and to configure a data storage module for the generative language model. The data storage module is used to store model dependency data. An application framework layer module is used to search for an application framework that matches the data storage module and the generative language model in the application framework layer based on the module selection information. The tools and operations support layer module is used to construct capability extension tool modules and operations support modules corresponding to the generative language model in the tools and operations support layer based on the module selection information. The data support layer module is used to construct the data preprocessing module and retrieval enhancement module of the generative language model in the data support and business application layer based on the module selection information; The business application layer module is used to construct the target application in the business application layer based on the generative language model, the data storage module, the application framework, the capability expansion tool module, the operation and maintenance support module, the data preprocessing module, and the retrieval enhancement module.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images