Industrial software middle table system based on large model

Abstract

The invention discloses an industrial software platform system based on a large model, and aims to improve the industrial data processing efficiency and the intelligent level by integrating a large model technology, domain knowledge and an information system interface. The system comprises a vertical large model module, an agent module and a universal API interface module. The vertical large model module adapts a general large model to a specific vertical field by using a GraphRAG technology and a LoRA fine tuning technology. The agent module comprises a memory module, a tool calling module and a planning module, and the autonomous decision-making ability of the large model is improved. The universal API interface module realizes seamless connection between a large model and an industrial information system, and opens up a gap between different information systems. According to the system, efficient data processing and intelligent decision making are achieved through knowledge graph construction, retrieval enhancement generation, model fine adjustment and a universal API interface, and the system adapts to market requirements and industrial environment changes.

Description

Technical Field

[0001] The present invention belongs to the field of industrial Internet and artificial intelligence, and specifically relates to technologies related to industrial software integrated big models, and in particular to an industrial software middle-office system based on big models. Background Art

[0002] With the development of society, the industrial Internet has been more and more widely used. Industrial enterprises are facing increasing demands for data processing and decision-making. Although traditional industrial information systems, such as ERP (Enterprise Resource Planning) and WMS (Warehouse Management System), have achieved information management to a certain extent, they have limitations in data processing and intelligent decision-making. These systems are often independent of each other, forming data islands.

[0003] In recent years, artificial intelligence technology, especially big model technology, has made significant progress. Through deep learning and massive data training, big models have powerful language understanding and generation capabilities. However, general big models have limitations in application in specific vertical fields and need to be enhanced and optimized through domain-specific knowledge bases and data sources. Summary of the invention

[0004] The present invention relates to an industrial software middle-end system based on a large model, which overcomes the limitations of existing industrial information systems and improves data processing efficiency and intelligence level by integrating large model technology, domain knowledge and information system interfaces.

[0005] The technical solution of the present invention is as follows:

[0006] An industrial software middle platform system based on a large model, comprising:

[0007] The vertical big model module encapsulates the general big model into a vertical big model suitable for vertical industry fields through retrieval enhancement generation and fine-tuning technology.

[0008] The intelligent agent module adds long-term memory, system tools, planning and other capabilities based on the vertical large model, further encapsulates it into an intelligent agent, and improves the ability of autonomous decision-making.

[0009] The general API interface module provides a series of general API interfaces, including general data input and output interfaces, system call interfaces, and result feedback call interfaces, for seamless connection between large models and industrial enterprise information systems.

[0010] in:

[0011] The vertical large model module uses GraphRAG technology to enhance the integration capability of domain knowledge. GraphRAG technology uses graph structure to enhance model generation. The specific steps are knowledge graph construction and retrieval enhancement generation.

[0012] The knowledge graph is used to structure domain knowledge into entities and relationships, extract knowledge points from the text using natural language processing technology, and construct a semantically consistent knowledge graph.

[0013] Among them, retrieval enhanced generation is used to dynamically retrieve relevant nodes and paths in the knowledge graph when generating answers in a large model, as generation conditions or prompt inputs, to enhance the accuracy of the answers.

[0014] Furthermore, the vertical large model module utilizes the LoRA fine-tuning technology to fine-tune the model by adding a low-rank layer to the model weights without changing the original weights of the model, so that the model can be quickly adapted to different fields and achieve a plug-and-play effect.

[0015] Furthermore, the intelligent agent module includes three parts: a memory module, a tool calling module, and a planning module.

[0016] The memory module combines long-term memory with short-term memory, records the interaction history of enterprise users and the history of task execution, supports context association, and dynamically adjusts the weight of memory content according to task requirements to ensure that the agent has efficient information acquisition capabilities during task execution. At the same time, key information is persistently stored to form a domain knowledge base to provide support for subsequent tasks.

[0017] The tool calling module provides a series of callable tools, such as network search, external API, etc., which support the agent to dynamically call according to task requirements. The agent can decompose complex tasks into several subtasks and combine tools to complete the efficient execution of the task chain.

[0018] The planning module uses the capabilities of the large model itself, and the agent can dynamically generate an execution path based on the task objectives. In multi-agent scenarios, it supports collaboration and resource sharing between agents, further improving the efficiency of task completion.

[0019] Furthermore, the general API interface module includes: data input and output general interface, system call general interface, and result feedback general interface.

[0020] The universal data input and output interface realizes unified access to structured data such as databases, semi-structured data such as JSON files, and unstructured data such as audio, pictures, and text, thereby realizing the collection, transmission, and processing of real-time data in enterprise information systems.

[0021] The system call interface is responsible for defining the interface format of the external information system, facilitating the interaction between the big data model and the external system. It also supports users to customize the parameters and processes of API calls according to specific business needs.

[0022] The result feedback calling interface obtains files in various output formats such as tables, charts, and texts returned by these systems according to the needs of different systems, and supports the collection and analysis of feedback data to continuously optimize the performance of large models.

[0023] The beneficial effects of the present invention are as follows:

[0024] The present invention provides a method for constructing an industrial software middle platform system based on a large model. Through GraphRAG and LoRA technology, the general large model is encapsulated into a vertical large model, and then further encapsulated into an intelligent body. By designing a general API interface module, seamless integration with the industrial enterprise information system is achieved. This not only improves the efficiency of industrial data processing and decision-making, but also enhances the flexibility and intelligence level of the system, enabling it to better adapt to rapidly changing market demands and complex industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a diagram of the system architecture of the industrial software platform based on the big model in the present invention;

[0026] Figure 2 This is an interface diagram of the industrial software middle platform system based on the big model in the present invention;

[0027] Figure 3 It is a schematic diagram of the process of the industrial software middle platform system based on the big model in the present invention; Specific implementation plan

[0028] The present invention will be further described below in conjunction with the accompanying drawings.

[0029] like Figure 1 As shown, the industrial software middle platform system based on a big model disclosed in the present invention consists of three main components: a vertical big model module, an intelligent body module and a general API interface module.

[0030] The vertical large model module includes the GraphRAG module and the model fine-tuning module.

[0031] The GraphRAG module builds a knowledge graph by integrating external texts. Its steps include the following core processes: First, extract key information from external texts, which usually relies on natural language processing techniques such as named entity recognition (NER) and relation extraction (RE) to identify entities and their relationships. The extracted information is converted into a structured data form, usually a triple (entity 1-relationship-entity 2). Then, these triples are used to create nodes and edges of the knowledge graph, where nodes represent entities and edges represent the relationships between them. In order to ensure the accuracy and semantic consistency of the knowledge graph, knowledge alignment and fusion techniques are used to merge or update the newly extracted information with the existing knowledge graph. In addition, the content in the graph is further transformed into a vector space representation through an embedding model for subsequent retrieval tasks. After the user enters the question, the system identifies the key information in it, such as entities, relationships, etc. Then, the system uses the structured information in the graph to find nodes and edges related to the query through the graph retrieval module, through graph traversal and similarity calculation, to find the subgraph that best matches the query. After obtaining the relevant subgraphs, generate answers or output results based on this information.

[0032] The model fine-tuning module uses the LoRA method to fine-tune the weight matrix in the pre-trained model by decomposing it into a low-rank matrix. In this process, the original weights are not directly modified, but the output of the model is adjusted by introducing a small number of trainable low-rank matrices. In specific operations, LoRA inserts low-rank adaptive parameter matrices in certain layers of the model, especially the self-attention layer and the feedforward network layer. These matrices are initialized to zero and gradually optimized with training. In the process of training using internal enterprise data, most of the parameters of the large model are frozen, and only the low-rank matrix introduced by LoRA is trained to improve training efficiency.

[0033] Furthermore, the agent module includes memory, tool, and planning modules.

[0034] The memory module implements long-term memory in the form of a dynamic knowledge base. It mediates between the input and output of the model and stores historical information in a long-term storage pool in a structured or unstructured form. Whenever the model encounters a new task or situation, it not only relies on the current parameters and context, but can also obtain previous information by querying the external memory base, thereby enhancing its ability to handle complex tasks. The memory module uses an event trigger mechanism to allow specific conversations, update the external memory storage based on new data, and decide which memories need to be stored or abandoned for a long time based on the memory priority mechanism.

[0035] The tool module provides a series of callable tools, such as web search, external API interface, database access, etc. These tools can be dynamically selected and called according to different task requirements. In this way, the agent can obtain and utilize external resources to improve the efficiency and accuracy of task execution. More importantly, the agent can break down complex tasks into several subtasks, and then call the appropriate tools to complete each subtask in turn, and finally summarize all the results to achieve the overall goal.

[0036] The planning module uses the computing and reasoning capabilities of the large model to help the agent dynamically generate the best execution path according to the task objectives. Through powerful modeling and reasoning capabilities, the agent can analyze the current task requirements, resource conditions, and environmental factors, and automatically select the most appropriate action strategy to ensure that the task is carried out efficiently and orderly. In the scenario of multi-agent collaboration, the planning module achieves the best task allocation by sharing task information, resource conditions, and execution progress in real time.

[0037] Further general API interface modules include data input and output general interface, system call general interface, and result feedback general interface.

[0038] The data input and output universal interface is used to process various types of data, including structured data such as information in databases, semi-structured data such as JSON files, and unstructured data such as audio, images, and text. This interface is used to clean and align data.

[0039] The system call general interface specifies the standard external system interface format. After the external system interface is encapsulated according to this format, when the system communicates with the large model, the large model can use function calls to obtain data from the system.

[0040] The result feedback calling interface focuses on obtaining output from various systems and supports multiple output formats, including tables, charts, and text.

[0041] Figure 2 It is the system interface of the industrial software platform based on the big model in the present invention.

[0042] The main interfaces of this system include large model dialogue, chat history, knowledge base, model fine-tuning, information system, tool management, and memory management interface.

[0043] In the big model dialogue interface, you can talk to the big model. The big model will decide whether to call the external system interface based on the user's input, and then generate graph-based retrieval enhancements based on the local database. After the answer is generated, the system determines whether the memory needs to be updated based on the event trigger mechanism.

[0044] In the chat history interface, you can check previous conversations with the big model.

[0045] In the knowledge base interface, you can import external data, supporting formats such as Excel, PDF, and Jpg. After the data is input and processed, the relevant knowledge graph will be automatically generated. In this interface, you can also preview the knowledge graph in 3D and add, delete, and modify nodes.

[0046] In the model fine-tuning interface, you can input data to fine-tune the model. The LoRA file generated after model fine-tuning will be saved in the file storage system, so that one-time training can be used multiple times. Users can select the required LoRA according to their needs in this interface.

[0047] In the information system interface, you can add external information system interfaces, which need to meet the general format defined in the system call general interface. After adding, the big model will determine when to call the system interface based on user input.

[0048] In the tool management interface, you can add various large model tools, such as online search. After adding, the large model will determine when to call the tool based on user input.

[0049] In the memory management interface, all key long-term memories sorted out from the conversation will be displayed. These memories can be added, deleted, and modified, and you can also add memory entries yourself.

[0050] like Figure 3 As shown in the figure, this is the flow chart of the system calling the large model to generate answers. After the user enters the dialogue, the intelligent agent first makes judgments and decomposes the tasks, generates a series of plans, and then gradually calls tools, external interfaces or retrieves knowledge bases and memories according to the plans, and finally generates answers. After the answer is generated, it is determined whether the memory needs to be updated for this question and answer. If necessary, the memory is updated, otherwise the process ends.

[0051] The above contents are further detailed descriptions of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can also make several substitutions or modifications to these described embodiments without departing from the concept of the present invention, and these substitutions or modifications should be regarded as belonging to the protection scope of the present invention.

[0052] Parts of the present invention that are not described in detail belong to the well-known technologies of those skilled in the art.

[0053] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial software middle platform system based on a large model, characterized in that: Including vertical large model module, intelligent agent module and general API interface module; Among them, the vertical large model module encapsulates the general large model into a vertical large model through RAG and fine-tuning technology; The agent module further encapsulates the vertical model into an agent by adding memory, tools, and planning modules; The general API interface module is used to connect to the enterprise information system to achieve seamless integration.

2. According to claim 1, an industrial software middle platform system based on a large model is characterized in that: The vertical large model module integrates domain-specific knowledge bases and data sources and uses RAG and fine-tuning technology to enhance the generation capability of general large models.

3. According to claim 2, an industrial software middle platform system based on a large model is characterized in that: The general API interface module includes a general interface for data input and output, a general interface for system calls, and a result feedback call interface to achieve effective communication and collaboration between the large model and the external system.

Images