An enterprise intelligent decision-making method and system based on a large model and a knowledge graph

Abstract

The application relates to the technical field of artificial intelligence, in particular to an enterprise intelligent decision-making method and system based on a large model and a knowledge graph, which comprises the following steps: constructing an enterprise global resource center, fusing global resources such as multi-modal data, an expert knowledge base and a business system and the like; building an enterprise global dynamic knowledge graph based on an ontology; constructing an AI large model training and reasoning engine, completing model training and reasoning optimization of a decision-making scene; and constructing a Multi-Agent collaborative intelligent application, realizing intelligent analysis and decision-making output of an enterprise in multiple scenes. Meanwhile, in view of the noise interference, high reasoning cost and poor result consistency of the existing knowledge graph large language model, a three-stage training optimization algorithm is proposed to realize directional optimization of the performance of the large model. The application breaks the barriers between data and business by fusing the advantages of the knowledge graph and the capabilities of the large model, constructs an enterprise intelligent closed loop of "perception-thinking-action", and significantly improves the intelligent and accurate levels of enterprise decision-making.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an enterprise intelligent decision-making method and system based on large models and knowledge graphs. Background Technology

[0002] With the rapid development of artificial intelligence technology and its deep integration with the digital economy, the complexity of the business environment and the intensity of market competition are intensifying simultaneously. Real-time and accurate decision-making capabilities have become a key support for building core competitiveness. The corporate decision-making process needs to comprehensively integrate internal and external resources across multiple dimensions, specifically encompassing internal business operation data and financial data, external market dynamics, policies and regulations, and information on industry competitors. Simultaneously, it needs to deeply incorporate the experience of domain experts to adapt to the decision-making needs of diverse business scenarios.

[0003] The current enterprise decision-making system generally suffers from a core problem: data fragmentation and a disconnect in the decision-making process. This manifests in three main pain points: First, the problem of data silos is prominent. Enterprise-wide data resources are scattered across dozens of business systems, exhibiting multi-source, heterogeneous, and multimodal characteristics. This results in inconsistent data formats, low correlation, and semantic ambiguity, leading to inefficient cross-departmental and cross-scenario data resource integration. Furthermore, the integration results often become invalid due to insufficient timeliness, creating a data gap in decision-making. Second, there is a gap between insight and execution. Traditional decision-making models heavily rely on human experience and judgment, leading to significant subjective biases. Moreover, analytical insights are difficult to directly translate into executable instructions for the operational system, hindering the effective implementation of decision-making value. Third, real-time decision support capabilities are insufficient. With the accelerating pace of market changes, traditional decision-making systems and technical solutions exhibit significant lag in data processing and analytical reasoning, making it difficult for enterprises to quickly seize market opportunities and effectively mitigate operational risks.

[0004] To address these pain points, artificial intelligence (AI) technology has become a core driving force for upgrading enterprise decision-making systems. The outstanding performance of Large Language Models (LLMs) in complex reasoning tasks represents a significant breakthrough in the field of AI. Reasoning ability (deriving logical conclusions from existing knowledge) is a core objective of AI, while knowledge graphs, representing relationships between entities in a structured form, are a crucial foundation for reasoning tasks. Theoretically, both can leverage their respective strengths: the generalized understanding capability of LLMs and the structured knowledge reasoning capability of knowledge graphs can break down data barriers and the disconnect between AI technology and business scenarios, constructing an intelligent decision-making closed loop of "perception-thinking-action," and transforming fragmented data into tangible decision-making value. However, in practical applications, current decision-making schemes based on knowledge graphs for large language models still have significant technical shortcomings: First, they are weak in resisting noise interference, lack efficient noise filtering and information purification mechanisms, making it difficult to accurately select effective decision information from multimodal data, which can easily lead to decision bias; second, there is an imbalance between reasoning efficiency and cost, as decision reasoning in complex business scenarios requires a large amount of knowledge matching calculations, which not only consumes huge amounts of computing resources but also has a significant lag in the reasoning process, failing to meet the needs of real-time decision-making; third, the consistency of reasoning results is poor, easily affected by factors such as the order of data input and the way it is expressed, resulting in large differences in output results, making it difficult to support large-scale and standardized enterprise decision-making applications.

[0005] In summary, existing technological solutions have failed to fundamentally address the core issue of data fragmentation and the disconnect between data and decision-making, making it difficult to meet enterprises' core needs for intelligent and precise decision-making and end-to-end business empowerment. Therefore, there is an urgent need for an intelligent enterprise decision-making method and system that can effectively break down data silos, connect the entire insight-execution chain, and support real-time decision-making. Summary of the Invention

[0006] The purpose of this invention is to provide an enterprise intelligent decision-making method and system based on large models and knowledge graphs to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, one aspect of the present invention provides an enterprise intelligent decision-making method based on large models and knowledge graphs, comprising the following steps: S1. Build an enterprise-wide resource platform: integrate and logically merge multimodal data assets, expert knowledge bases, and all data, knowledge, and system resources carried by business systems. S2. Construct an ontology-driven enterprise global dynamic knowledge graph: Using the enterprise full-domain resource platform as the core data source and knowledge source, the core knowledge elements required for enterprise decision-making are obtained through multi-dimensional extraction and modeling of entities, relationships, attributes and business actions. S3. Building an AI large-scale model training and inference engine: Relying on a knowledge graph-based three-stage training engine and hybrid inference engine for large models, combined with high-performance computing, model fine-tuning technology, and preference alignment, the training and inference of large AI models are completed. S4. Construct intelligent applications based on a multi-Agent collaborative architecture: Integrate and deploy multiple functional components such as intelligent data analysis agent, intelligent report generation agent, decision generation and execution agent, and feedback learning agent. Each agent is implemented based on a hybrid agent architecture that integrates the ReAct framework and the Workflow mode. Each agent realizes intelligent analysis of enterprise business data, automatic generation of decision reports, intelligent output of decision solutions, intelligent execution, and collection of feedback data by calling the training and inference results of large AI models.

[0008] Preferably, the construction of the enterprise-wide resource platform in step S1 specifically includes: A unified semantic mapping and ontology mapping engine enables semantic alignment and ontology association across data sources and system resources; data virtual mapping and logical orchestration technologies are used to standardize and encapsulate various heterogeneous resources into reusable and scalable core resource units, which are then incorporated into the unified management of the enterprise's full-domain resource platform, forming a standardized, ontological, and service-oriented core resource set; The multimodal data assets include data warehouses, data lakes, business system databases, enterprise-related documents, images, audio, and video. The expert knowledge base includes the enterprise's private domain knowledge base, business knowledge logic, enterprise operating indicators, and historical decision-making cases; The business systems include MES system, OMS system, CRM system, OA system, ERP system, and BI system; The unified semantic mapping and ontology mapping engine constructs an enterprise-level unified data dictionary and ontology vocabulary based on the inherent attributes of resources and business characteristics.

[0009] Preferably, the construction of an ontology-driven enterprise global dynamic knowledge graph in step S2 specifically includes: After multiple processing operations such as entity disambiguation, cross-source alignment, knowledge fusion and ontology mapping, an enterprise global knowledge graph containing real-time dynamic knowledge is constructed and stored in a graph database to realize the structured organization, dynamic updating and efficient retrieval of knowledge. The core modeling system of the enterprise global dynamic knowledge graph includes an entity type system, a relationship type system, an entity attribute system, and a business action strategy system. The entity type system includes core enterprise entities, business process entities, product entities, customer entities, and hierarchical entity structures. The relationship type system includes one-to-one, one-to-many, and many-to-many association patterns and semantic constraints; The entity attribute system includes basic attributes, business attributes, and status attributes; The business action strategy system includes association rules, algorithm programs, and system API interfaces; Entity extraction and relation extraction are implemented using a deep learning model based on an attention mechanism.

[0010] Preferably, the construction of the AI ​​large model training and inference engine in step S3 specifically includes: The large model three-stage training engine employs a specifically designed large model three-stage training algorithm, which includes: Narrowing the knowledge graph retrieval scope: Based on the predefined entity type and relation type matching rules in the enterprise global dynamic knowledge graph, knowledge graph subgraphs related to the target reasoning task are selected. Specifically, this includes obtaining the text description of the target reasoning task, parsing the core entities and relationships of the task, comparing them with the entity type and relation type matching rules of the knowledge graph, extracting knowledge graph subgraphs containing the core entities and relationships based on efficient graph database retrieval algorithms, and automatically filtering irrelevant nodes and invalid relation links; finally, embedding the knowledge graph subgraphs into the corresponding graph representation using a graph neural network. Screening high-quality paths in the knowledge graph: For the reasoning paths in the subgraphs of the knowledge graph, a dual-dimensional evaluation of relevance and completeness is conducted. The relevance evaluation is quantitatively represented by calculating the semantic association weight between the reasoning path and the target task, while the completeness evaluation is completed by determining the degree to which the reasoning path covers the core elements of the target task. By fusing the relevance and completeness indicators of the paths, a path evaluation function is constructed using a neural network model. Based on the evaluation results, high-quality knowledge paths that meet the requirements of the target reasoning task are ranked and screened to reduce interference from redundant paths. The large model is trained and optimized in two steps: First, supervised fine-tuning is performed on the basic AI model using sample data corresponding to the high-quality knowledge paths. The sample data includes the input question, knowledge graph path, and standard output results. By minimizing the objective function, the model's adaptability to knowledge graph information is improved. Second, the preference alignment algorithm is used to optimize the model after SFT fine-tuning. By using manually labeled preference sample sets and optimizing the model's reward function, the model output is made to better meet the needs of enterprise decision-making.

[0011] Preferably, the node representation update logic of the graph neural network is as follows: at the first... Layers, nodes The representation is updated as follows: ; in, Represents a node In the Layer representation; The relation matrix representing the learning process; Indicates aggregation operation; Indicates an update operation; initial representation and Random initialization, The union of the neighborhoods of all subject entities. For son Figure 3 A set of tuples; r represents the graph relationship between node e' and node e.

[0012] Preferably, the path evaluation function is defined as follows: ; MLP stands for Multilayer Perceptron; Indicates the activation function; This represents the input query text; Indicates the cumulative value of the path; Estimate the value of the path; Among them, the cumulative value of the path The calculation method is as follows: ; in, Indicates from arrive The path set; Represents a node Update logic, , Represents a node The total number of floors; Path Prediction Value The calculation method is as follows: ; in, It is a feedforward neural network. This indicates concatenating two vectors; The path evaluation function is trained using the cross-entropy loss function, which is defined as follows: ; Wherein represents the query The correct set of entities; This indicates the set of other entities in the query subgraph that are incorrect.

[0013] Preferably, the objective function of the supervised fine-tuning is: ; in, These represent the model parameters of a basic AI model. Represented as The key reasoning path for screening; The parameter is AI large models in known and Time generation The conditional probability; This represents the training dataset; This represents the expectation calculation for the training samples; The objective function for preference alignment is: ; in, This indicates a preference for aligning datasets; Indicates the most important preferred path; This represents a less important secondary path randomly sampled in the subgraph; This represents the temperature parameter, used to control the degree to which the model deviates from the reference model; This represents the sigmoid function.

[0014] Preferably, step S3 further includes: The hybrid reasoning engine adopts a combination of knowledge graph symbolic reasoning and large-model semantic reasoning. The high-performance computing relies on GPU clusters to support parallel acceleration of model training and inference.

[0015] Preferably, step S3, which involves constructing an intelligent application based on a Multi-Agent collaborative architecture, specifically includes: The Hybrid Agent architecture is adopted, and the execution logic is as follows: Complex tasks are broken down and executed by B-Agent based on the ReAct framework, while simple tasks are quickly responded to by S-Agent based on the Workflow pattern. By leveraging feedback learning agents to build a closed-loop feedback iteration mechanism of demand analysis, task execution, and reflection iteration, the performance of large AI models can be continuously optimized, supporting enterprises in real-time insight, intelligent decision-making, and business empowerment.

[0016] Another aspect of the present invention provides an enterprise intelligent decision-making system based on large models and knowledge graphs, for implementing the enterprise intelligent decision-making method based on large models and knowledge graphs as described above, including: The enterprise-wide resource platform construction module is used to integrate and logically fuse multimodal data assets, expert knowledge bases, and various data, knowledge, and system resources corresponding to enterprise business systems. It achieves semantic alignment and ontology association across data sources and system resources through a unified semantic mapping and ontology mapping engine. It adopts data virtual mapping and logical orchestration technology to standardize and encapsulate various heterogeneous resources into reusable and scalable core resource units, which are then incorporated into the unified management of the enterprise-wide resource platform to form a standardized, ontological, and service-oriented core resource set. The Enterprise Global Dynamic Knowledge Graph Construction Module is used to build an ontology-based enterprise global dynamic knowledge graph. Specifically, it includes: an element extraction unit, which extracts core elements needed for enterprise decision-making through multi-dimensional extraction of entities, relationships, attributes, and business actions; a fusion and alignment unit, which performs entity disambiguation and cross-source alignment on the extracted core element information; and a graph construction and storage unit, which, after knowledge fusion and ontology mapping, constructs an enterprise global knowledge graph containing real-time dynamic knowledge and stores it in a graph database, achieving structured organization, dynamic updates, and efficient retrieval of knowledge. The AI ​​large-scale model training and inference framework module is used to build the training and inference framework and complete the training and inference of the AI ​​large-scale model. Specifically, it includes: a three-stage training engine unit, which is configured with three-stage training algorithm logic, including logic for narrowing the search scope, logic for selecting high-quality paths, and logic for two-step training optimization; a hybrid inference engine unit, which adopts inference logic that combines knowledge graph symbolic inference and large-scale model semantic inference; a high-performance computing unit, which relies on GPU clusters to provide parallel acceleration support; and a model fine-tuning unit, which is used to realize supervised fine-tuning and preference alignment optimization of the large-scale model. The multi-Agent collaborative intelligent application module is configured with an intelligent data analysis agent, an intelligent report generation agent, a decision inference and execution agent, and a feedback learning agent. Each agent is implemented based on a hybrid agent architecture that integrates the ReAct framework and the Workflow mode. The intelligent application module is used to call the inference results of the AI ​​large model training inference module to realize intelligent analysis of enterprise business data, automatic generation of decision reports, intelligent output of decision solutions, intelligent execution, and collection of feedback data.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an enterprise intelligent decision-making method and system based on large models and knowledge graphs. By constructing an enterprise-wide resource platform, it integrates multimodal data, expert knowledge bases, and business systems, achieving unified management and aggregation of multi-source heterogeneous resources. This provides comprehensive and accurate data and knowledge support for enterprise intelligent decision-making. Relying on ontology technology, it builds a global dynamic knowledge graph for the enterprise, systematically sorting out, modeling, and updating the core elements of enterprise decision-making in real time, thus constructing a knowledge network foundation for enterprise decision-making. Addressing key pain points of existing knowledge graph-enhanced large language models, such as noise interference, high inference costs, and poor result consistency, it innovatively designs a three-stage training optimization algorithm and constructs an AI large model training and inference engine. This enables targeted optimization of model performance in decision-making scenarios, significantly improving model training efficiency and inference quality. Finally, it constructs a multi-agent collaborative intelligent application module, empowering enterprises to automate the entire process of intelligent analysis and decision-making across multiple scenarios. This invention deeply integrates the advantages of structured knowledge representation of knowledge graphs with the general reasoning capabilities of large models, breaks down information barriers between enterprise data and business, and constructs a closed loop of intelligent enterprise decision-making of "perception-thinking-action". It effectively improves the real-time, accuracy, scientificity and intelligence of enterprise decision-making, and provides core technical support for empowering enterprise business and high-quality development. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of an enterprise intelligent decision-making method based on large models and knowledge graphs in an embodiment of the present invention. Figure 2 This is a flowchart of the three-stage training algorithm for the large model in an embodiment of the present invention; Figure 3 This is a flowchart of the Multi-Agent collaborative intelligent application module in an embodiment of the present invention; Figure 4 This is a structural block diagram of an enterprise intelligent decision-making system based on a large model and knowledge graph, as described in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the embodiments of the present invention provide an enterprise intelligent decision-making method based on large models and knowledge graphs, including the following steps: Before executing the following methods and steps, preliminary preparations must be completed: First, clearly define the target company's business domain (this implementation uses a manufacturing company as an example) and core decision-making scenarios (such as production scheduling optimization, supply chain risk warning, and precise customer operation), and delineate the data scope and business analysis boundaries accordingly; Second, build the basic hardware environment, including deploying a CPU cluster (using Intel Xeon Gold 6330 processors, a total of 32 nodes) for data preprocessing and knowledge graph construction, deploying a GPU cluster (using NVIDIA A100 GPUs, a total of 16 nodes, with 80GB of video memory per node) for large model training and inference, and configuring a distributed storage system (total storage capacity of 100TB, supporting PB-level expansion) for persistent storage of the company's core resource collection and knowledge graph data; Finally, build the software support environment, using CentOS 7.9 as the operating system, Spark 3.3.0 as the data processing framework, Neo4j 5.10 as the graph database (supporting efficient querying and transaction processing of massive graph data), and PyTorch 2.0 as the AI ​​large model training framework, ensuring compatibility and data interaction efficiency between modules; S100. Building an Enterprise-Wide Resource Platform: This platform integrates and logically merges multimodal data assets, expert knowledge bases, and all data, knowledge, and system resources carried by business systems. It achieves semantic alignment and ontology association across data sources and systems through a unified semantic mapping and ontology mapping engine. Using data virtual mapping and logical orchestration technologies, it standardizes and encapsulates various heterogeneous resources into reusable and scalable core resource units, incorporating them into the unified management of the enterprise platform to form a standardized, ontological, and service-oriented core resource set. S101, Data Access: Business system access is based on multi-source connectors and a query federation mechanism to achieve full access to heterogeneous resources. Data from enterprise business systems is accessed through API calls and query subscriptions. Specific access methods are as follows: 1) MES System: Real-time querying via OPC UA protocol, focusing on maintaining the query chain for core data such as production work orders, equipment operating parameters (e.g., machine tool speed, load current, temperature), and production cycle time; 2) Other business systems (OMS, CRM, OA, ERP (using SAPS / 4HANA), BI): Querying via JDBC interface, covering order information, customer files, contract data, financial data, and operating reports; Multimodal data assets and expert knowledge base access are achieved using a combination of "distributed query + local file index mapping". The expert knowledge base is collected and organized using structured methods to build a query index system. Specific access solutions are as follows: 1) Document data (PDF, Word, Excel): via Python... 1) Build query indexes for the enterprise's internal knowledge base and external industry reporting platforms, triggering content queries on demand; 2) Image data (production site monitoring images, product quality inspection images): Industrial cameras collect data in real time and store it on a local file server, only accessing the middle platform with file indexes, storage paths, and other metadata to maintain the query link; 3) Audio / video data (conference recordings, training videos): Accessed through the query interface of the enterprise's internal storage system, retaining the original storage location and maintaining on-demand query capabilities; 4) Expert knowledge base: Collected and organized through a combination of semi-structured interviews and structured questionnaires, building a query index system that covers the query relationships of the enterprise's private knowledge base, business knowledge logic, enterprise operating indicators, and historical decision-making cases; S102. Classification and Quality Control: A two-level classification method is adopted. The first-level classification is divided into three categories: "Data Asset Query Index," "Knowledge Asset Query Index," and "System Resource Query Index." The second-level classification is further refined (12 subcategories of data assets: production / sales / finance / customer / equipment data, etc.; 8 subcategories of knowledge assets: business process specifications / expert experience summaries / historical decision-making cases / industry standards and regulations, etc.; 6 subcategories of system resources: business system interfaces / algorithm model components / hardware device nodes, etc.). A "rule + machine learning" combination mode is adopted. First, a preliminary classification is performed using preset rules (file extensions, query field names, system identifiers, etc.), and then a Naive Bayes classification model is used for secondary classification of the fuzzy index, with an accuracy of ≥98.5%. Query link health rules are set to automatically detect and alert on the availability, response time, and data freshness of the classified indexes, ensuring the quality of query services. S103. Semantic Alignment and Query Association Construction: Based on ontology design principles, implement semantic alignment and query association across data sources and system resources to solve the "query semantic silo" problem: 1) Ontology vocabulary construction: Following the principles of "decision-oriented, lightweight, and object-oriented", focus on the core business entities and relationships required by the knowledge graph, and define core concepts with unified semantics and unique RIDs (resource identifiers); 2) Semantic mapping relationship establishment: Build a global semantic dictionary and a unified query dictionary to realize cross-system query field mapping (such as ERP "customer code", CRM "partner ID", OA "business unit number" are unified as "customer unique identifier"); 3) Association mechanism: Cover the semantic definition, query format, and association relationship of 800+ core query fields.

[0021] S200. Construct an ontology-driven enterprise global dynamic knowledge graph: Using the enterprise global resource platform as the core data source and knowledge source, the core knowledge elements required for enterprise decision-making are obtained through multi-dimensional extraction and modeling of entities, relationships, attributes and business actions; after processing such as entity disambiguation, cross-source alignment, knowledge fusion and ontology mapping, an enterprise global knowledge graph containing real-time dynamic knowledge is constructed and stored in a graph database to realize the structured organization, dynamic updating and efficient retrieval of knowledge; S201. The ontology architecture of the enterprise global dynamic knowledge graph adopts a three-level structure of "top-level ontology - domain ontology - task ontology". The top-level ontology defines general concepts (such as entities, relations, and attributes), the domain ontology defines core business concepts for manufacturing enterprises (such as production, sales, supply chain, and customers), and the task ontology defines special concepts for specific decision-making scenarios (such as "work order priority" and "equipment utilization" in production scheduling optimization). S202, Full Data Acquisition: Internal data comes from the enterprise core resource set built by S100, covering data across the entire business chain, including production, sales, finance, and customers; external data is collected through legal channels, including industry policies and regulations, industry dynamics, competitor data, and data from upstream and downstream enterprises in the supply chain. S203. Core Knowledge Element Extraction: A deep learning model based on an attention mechanism (BERT-BiLSTM-CRF) is used to complete entity extraction, relation extraction, and attribute extraction. The model input is text data that has undergone word segmentation (jieba segmentation), part-of-speech tagging (HanLP tool), and entity type pre-annotation. A training dataset is constructed by randomly selecting 100,000 text data from the enterprise's core resource set. Three professional annotators perform entity, relation, and attribute annotation, resulting in 80,000 valid training samples. The training samples are divided into training and validation sets in an 8:2 ratio. The model training parameters are set as follows: batch size of 32, learning rate of 2e-5, number of training epochs of 15, and the optimizer is AdamW. After training, the model's entity extraction accuracy reaches 92.3%, relation extraction accuracy reaches 89.7%, and attribute extraction accuracy reaches 91.5%. S204. The extracted core knowledge elements follow four multi-dimensional modeling systems: the entity type system covers 12 major categories of entities, including core enterprise entities, business process entities, product entities, customer entities, and hierarchical entity structures, which are further subdivided into 50+ subcategories; the relationship type system includes one-to-one, one-to-many, and many-to-many association patterns and semantic constraints, defining 30+ core relationship types; the entity attribute system involves basic attributes, business attributes, and state attributes, with each entity containing 10-30 core attributes; and the action strategy system includes association rules, algorithm programs, and system API interfaces. S205, Entity Disambiguation and Cross-Source Alignment: To address the issue of homonymous and heteronymous entities during the extraction process, a "rule matching + semantic similarity calculation" approach is adopted for disambiguation: First, a rule base is established, which includes the standard entity names, aliases, and abbreviations defined within the enterprise. Preliminary disambiguation is completed through exact matching. For entities not covered by the rule base, the cosine similarity algorithm is used to calculate the semantic vector similarity of the entity description text (the semantic vector is generated by the BERT model). The similarity threshold is set to 0.85. When the semantic similarity of two entities is ≥0.85, they are determined to be the same entity. S206. The alignment and integration process is divided into internal integration and external integration: Internal integration aligns the element information extracted from different business systems of the enterprise, using the entity code of the ERP system as the benchmark to unify the entity identifiers of each system; External integration aligns the element information extracted from external data with the internal element information, completing the alignment through the combination matching method of "entity name + business attribute", and finally forming a unified set of core element information; Knowledge integration is carried out to integrate the attribute information of the same entity in multi-source data, such as merging customer revenue contribution data obtained from ERP and BI systems; S207. Knowledge Graph Construction and Storage: Based on the ontology architecture and the core element information after alignment and fusion, a global dynamic knowledge graph for the enterprise is constructed using the Neo4j graph database. Entities, relationships, attributes, and action strategies are defined using Cypher statements. A dynamic update mechanism for the knowledge graph is established, achieving real-time updates through incremental data collection and extraction. Each update only collects data added or changed within the previous cycle. Incremental core element information is obtained through extraction, disambiguation, and alignment processes and integrated into the existing graph. The knowledge graph constructed in this implementation contains 125,000 entity nodes, 358,000 relationship edges, and 860 action strategy nodes, and is stored in the Neo4j graph database.

[0022] S300, Building an AI Large-Scale Model Training and Inference Engine: Relying on a knowledge graph-based three-stage training engine and hybrid inference engine for large models, combining high-performance computing, model fine-tuning techniques, and preference alignment to complete the training and inference of large models; among which, such as Figure 2 As shown, the large model three-stage training engine adopts a specially designed large model three-stage training algorithm, which solves the problems of "high noise, high cost and poor consistency" of existing knowledge graph-based large language model reasoning methods. S301. Basic Environment Configuration: Deploy the PyTorch 2.0 training framework on the GPU cluster nodes, configure the CUDA 11.7 deep learning computing library, adopt the distributed training strategy (DDP) to achieve multi-GPU parallel training, and optimize the data transmission efficiency between nodes through the NCCL communication library; build a model caching mechanism to cache the pre-trained model parameters to GPU video memory and system memory to reduce model loading time; this implementation method selects the open-source Qwen3-235B-A22B model as the basic large model; The S302 hybrid inference engine combines knowledge graph symbolic reasoning with large-scale model semantic reasoning: for deterministic rule-based reasoning, it directly matches association rules through knowledge graph symbolic reasoning; for complex prediction-based reasoning, it generates results through large-scale model semantic reasoning and verifies the rationality of the reasoning logic by combining knowledge graph paths; high-performance computing is achieved through GPU clusters, and TensorRT is used to optimize the inference model, reducing model inference latency and improving training speed by 16 times compared to single GPU training. S310. Narrow the scope of knowledge graph retrieval. Based on the predefined entity type and relation type matching rules in the enterprise global dynamic knowledge graph, filter out knowledge graph subgraphs related to the target reasoning task, remove irrelevant knowledge nodes and relation links, and reduce noise data input. S311. Taking the "production scheduling optimization" decision task as an example, obtain the text description of the target reasoning task: "Based on the current operating status of each production line equipment, production work order progress, and raw material inventory, optimize the production scheduling plan for the next 24 hours to ensure that the on-time delivery rate of orders is over 90% and the equipment utilization rate is over 85%." S312. Analysis of core elements of the task: The core entities are "production line", "equipment", "production work order" and "raw material inventory", and the relationships are "equipment-operating status", "production work order-progress" and "raw material inventory-production line". S313, Determining retrieval constraints: The core entities and relationships obtained from parsing are compared with the entity type and relationship type matching rules of the enterprise's global dynamic knowledge graph to determine the retrieval constraints as "knowledge subgraph containing core entities and corresponding relationships"; S314. Knowledge Graph Subgraph Extraction: Based on the Neo4j graph database, the Breadth-First Search (BFS) algorithm for breadth-first search is used with a search depth of 3 layers to extract knowledge graph subgraphs containing core entities and relationships. Irrelevant nodes and invalid relationship links are automatically filtered out. The final subgraph contains 860 entity nodes and 2150 relationship edges, which is 92% less than the total knowledge graph data volume. S315. Use graph neural networks to embed subgraphs into their corresponding graph representations, and set the number of model layers. The aggregation function AGG uses MeanPooling, and the update function UPDATE uses the ReLU activation function; initial representation Initialized using Word2Vec pre-trained vectors (1600 dimensions), and with random initialization of the relation matrix, the node representation update logic of the graph neural network is as follows: At the... Layers, nodes The representation is updated as follows: ; in, Represents a node In the Layer representation; The relation matrix representing the learning process; Indicates aggregation operation; Indicates an update operation; initial representation and Random initialization, The union of the neighborhoods of all subject entities. For son Figure 3 A set of tuples; r represents the graph relationship between node e' and node e; S320. Screen high-quality paths in the knowledge graph. For reasoning paths in the subgraphs of the knowledge graph, conduct a two-dimensional evaluation of relevance and completeness to screen high-quality knowledge paths that meet the requirements of the target reasoning task and reduce the interference of redundant paths.

[0023] S321. Reasoning Path Generation: Using a path search algorithm, all reasoning paths from the core entity to the decision target in the subgraph are generated, resulting in a total of 42 initial reasoning paths.

[0024] S322. Relevance Assessment: Calculating the Cumulative Value of Pathways That is, the sum of the vector products of each entity-relation pair in the path and the transpose of the query vector (the BERT vector of the target task), as shown in the formula: ; in, Indicates from node To the node The path set; Represents a node Update logic, , Represents a node The total number of floors; S323. Integrity Assessment: Calculate the estimated value of the path. It is calculated by a feedforward neural network, and the formula is: ; in, It is a feedforward neural network. This indicates concatenating two vectors; S324. Construct the path evaluation function, the formula is: ; MLP stands for Multilayer Perceptron; Indicates the activation function; This represents the input query text; Indicates the cumulative value of the path; Estimate the value of the path; Then, the cross-entropy loss function is used to train and evaluate the model. The training dataset contains 1000 labeled samples. After 100 iterations, the model converges (loss value ≤ 0.05). The cross-entropy loss function is defined as follows: ; in, Indicates query The correct set of entities; This represents the set of other entities in the query subgraph that are incorrect. S325. Based on the total value of the paths, 12 high-quality knowledge paths were finally selected, such as "raw material inventory → production line → equipment → operating status → production work order → progress → production scheduling plan", which effectively reduced the interference of redundant paths. S330, a two-step training optimization model, improves the model's ability to adapt to structured information and the consistency and accuracy of inference results through supervised fine-tuning and preference alignment algorithms; S331. The first step is to perform supervised fine-tuning (SFT): Based on the selected high-quality knowledge paths, construct a structured training dataset. It contains 10,000 samples, each in the format of "input question - knowledge graph path hint - standard output result"; the training samples are divided into training and validation sets in a 9:1 ratio, and the Qwen3-235B-A22B basic model is fine-tuned. The training parameters are set as follows: batch size of 16, learning rate of 1e-5, number of training epochs of 10, and AdamW optimizer is used; the optimization objective function is: ; in, These represent the model parameters of a basic AI model. Represented as The key reasoning path for screening; The parameter is AI large models in known and Time generation The conditional probability; This represents the training dataset; This represents the expected value calculation for the training samples; The trained model achieved an accuracy of 88.6% on the validation set. S332, Step Two: Perform Preference Alignment (DPO) Optimization: Construct a DPO Preference Sample Set It contains 5000 manually labeled samples, each sample containing both high-quality and low-quality outputs corresponding to the same input; the temperature parameter β is set to 0.9, and it is trained using the PyTorch framework with 5 training epochs; the optimization objective function is: ; in, This indicates a preference for aligning datasets; Indicates the most important preferred path; This represents a less important secondary path randomly sampled in the subgraph; This represents the temperature parameter, used to control the degree to which the model deviates from the reference model; Represents the sigmoid function; The optimized model achieved a 95.3% rate of high-quality outputs. S340. Large model inference based on knowledge graph relies on high-performance computing and GPU clusters. The GPU cluster adopts a multi-node heterogeneous collaborative computing mode. Through task sharding and data parallelism strategies, the large model inference task is decomposed into multiple independent sub-tasks and allocated to different GPU nodes for synchronous execution. At the same time, the inference calculation process is optimized by combining the structured features of knowledge graphs, which supports parallel acceleration of model inference. S400, intelligent applications based on a multi-agent collaborative architecture: Integrates and deploys various functional agents. Each agent is implemented based on a hybrid agent architecture that combines the ReAct framework and Workflow mode. It calls upon the inference results of the AI ​​large-scale model training inference module to enable intelligent decision-making empowerment across all enterprise business scenarios, such as... Figure 3 As shown; S401 The intelligent application module is deployed using a microservice architecture. Each Agent is an independent microservice component and interacts with data through message queues. The intelligent data analysis Agent is deployed on a CPU cluster and uses the streaming computing framework Flink to achieve real-time data processing. The analysis dimensions include multiple dimensions such as production, supply chain, and customers, and the analysis results are output in a structured manner. S402 Intelligent Report Generation Agent automatically generates decision reports based on data analysis results and large model inference conclusions, and supports export in Word, PDF and HTML formats. The report generation adopts a template-based + dynamic filling method, with multiple preset report templates. It obtains inference conclusions and completes the filling by calling the large model API interface. The generation process does not require manual intervention. S403, Decision Generation and Execution Agent is responsible for generating and tracking the execution of decision-making schemes. The generated decision-making schemes are executed by calling enterprise business systems through API interfaces (such as calling the MES system to adjust production work orders, calling the ERP system to trigger procurement processes); it also supports multi-scenario decision simulation, real-time tracking of decision execution progress, and collection of execution process data. S404. The feedback learning agent collects actual data and result feedback after decision execution, constructs a feedback sample set, and the sample format is "decision input-decision plan-execution result-evaluation score". The evaluation score is labeled by business personnel according to the completion of decision objectives. The feedback sample set is used as incremental training data and is periodically input into the AI ​​large model to train the inference module to achieve continuous optimization of the model. The execution logic of the S405 and Hybrid Agent hybrid architecture is as follows: Users initiate requests through the enterprise's internal intelligent decision-making platform (supporting text / voice interaction), and the multi-turn dialogue management module and intent recognition module parse the core requirements; complex tasks are decomposed into sub-tasks by the B-Agent based on the ReAct framework and assigned to the corresponding functional Agents for parallel / serial execution; simple tasks are quickly responded to by the S-Agent based on the Workflow mode. S500 System Deployment and Testing Verification: S501. The enterprise intelligent decision-making method of the present invention adopts a microservice architecture and is deployed on a Kubernetes cluster. Each module (enterprise global resource platform, knowledge graph construction module, large model training and inference module, and intelligent application module) is deployed independently and elastically scaled; Redis is used to cache hot data to improve system response speed; Nginx is used to achieve load balancing and support 1000 concurrent users. S502: Functional testing verified the accuracy of data integration, the completeness of knowledge graph construction, and the accuracy of model inference. All functional modules met the design requirements. Performance testing showed that the model training efficiency was greatly improved compared with traditional methods, and the concurrent processing capability supported 1000 users operating simultaneously.

[0025] like Figure 4 As shown, this invention also provides an enterprise intelligent decision-making system based on large models and knowledge graphs, used to implement the enterprise intelligent decision-making method based on large models and knowledge graphs as described above, including: The Enterprise-wide Resource Platform Module is used to integrate and logically fuse multimodal data assets, expert knowledge bases, and various data, knowledge, and system resources corresponding to enterprise business systems. Deployed on a distributed server cluster, it adopts the Spring Cloud microservice architecture and includes resource access submodules, classification processing submodules, standardization submodules, and storage submodules. It achieves semantic alignment and ontology association across data sources and system resources through a unified semantic mapping and ontology mapping engine. Using data virtual mapping and logical orchestration technology, it standardizes and encapsulates various heterogeneous resources into reusable and scalable core resource units, which are then incorporated into the unified management of the Enterprise-wide Resource Platform to form a standardized, ontological, and service-oriented core resource set supporting 1000 concurrent requests. The Enterprise Global Dynamic Knowledge Graph Construction Module is used to build an ontology-based enterprise global dynamic knowledge graph. Deployed on a CPU cluster and developed in Python, it includes: an element extraction unit, which extracts core elements needed for enterprise decision-making through multi-dimensional extraction of entities, relationships, attributes, and business actions; a fusion and alignment unit, which performs entity disambiguation and cross-source alignment on the extracted core element information; and a graph construction and storage unit, which constructs an enterprise global knowledge graph containing real-time dynamic knowledge after knowledge fusion and ontology mapping, and stores it in a graph database to achieve structured organization, dynamic updates, and efficient retrieval of knowledge. The AI ​​large-scale model training and inference module is deployed on a GPU cluster and developed using the PyTorch framework. It is used to build the training and inference framework and complete the training and inference of large AI models. Specifically, it includes: a three-stage training engine unit with three-stage training algorithm logic, including logic for narrowing the search scope, logic for selecting high-quality paths, and two-step training optimization logic; a hybrid inference engine unit that combines knowledge graph symbolic inference with large-scale model semantic inference; a high-performance computing unit that leverages the GPU cluster for parallel acceleration support; and a model fine-tuning unit for supervised fine-tuning and preference alignment optimization of large models. Its core function is to complete the training and inference of large models, achieving a 16-fold improvement in model training efficiency compared to a single GPU. The multi-Agent collaborative intelligent application module is deployed on a cloud server and developed using a front-end and back-end separation architecture. It is configured with an intelligent data analysis agent, an intelligent report generation agent, a decision inference and execution agent, and a feedback learning agent. Each agent is implemented based on a hybrid agent architecture that integrates the ReAct framework and the Workflow mode. The intelligent application module is used to call the inference results of the AI ​​large model training inference module to realize intelligent analysis of enterprise business data, automatic generation of decision reports, intelligent output of decision solutions, intelligent execution, and collection of feedback data, thereby enabling intelligent empowerment of the entire decision-making process and supporting multiple users to operate online simultaneously.

[0026] This invention provides an enterprise intelligent decision-making method and system based on large models and knowledge graphs. By constructing an enterprise-wide resource platform, it integrates multimodal data, expert knowledge bases, and business systems, achieving unified management and aggregation of multi-source heterogeneous resources. This provides comprehensive and accurate data and knowledge support for enterprise intelligent decision-making. Relying on ontology technology, it builds a global dynamic knowledge graph for the enterprise, systematically sorting out, modeling, and updating the core elements of enterprise decision-making in real time, thus constructing a knowledge network foundation for enterprise decision-making. Addressing key pain points of existing knowledge graph-enhanced large language models, such as noise interference, high inference costs, and poor result consistency, it innovatively designs a three-stage training optimization algorithm and constructs an AI large model training and inference engine. This enables targeted optimization of model performance in decision-making scenarios, significantly improving model training efficiency and inference quality. Finally, it constructs a multi-agent collaborative intelligent application module, empowering enterprises to automate the entire process of intelligent analysis and decision-making across multiple scenarios. This invention deeply integrates the advantages of structured knowledge representation of knowledge graphs with the general reasoning capabilities of large models, breaks down information barriers between enterprise data and business, and constructs a closed loop of intelligent enterprise decision-making of "perception-thinking-action". It effectively improves the real-time, accuracy, scientificity and intelligence of enterprise decision-making, and provides core technical support for empowering enterprise business and high-quality development.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An enterprise intelligent decision-making method based on large models and knowledge graphs, characterized in that, The method includes the following steps: S1. Build an enterprise-wide resource platform: integrate and logically merge multimodal data assets, expert knowledge bases, and all data, knowledge, and system resources carried by business systems. S2. Construct an ontology-driven enterprise global dynamic knowledge graph: Using the enterprise full-domain resource platform as the core data source and knowledge source, the core knowledge elements required for enterprise decision-making are obtained through multi-dimensional extraction and modeling of entities, relationships, attributes and business actions. S3. Building an AI large-scale model training and inference engine: Relying on a knowledge graph-based three-stage training engine and hybrid inference engine for large models, combined with high-performance computing, model fine-tuning technology, and preference alignment, the training and inference of large AI models are completed. S4. Construct intelligent applications based on a multi-Agent collaborative architecture: Integrate and deploy multiple functional components such as intelligent data analysis agent, intelligent report generation agent, decision generation and execution agent, and feedback learning agent. Each agent is implemented based on a hybrid agent architecture that integrates the ReAct framework and the Workflow mode. Each agent realizes intelligent analysis of enterprise business data, automatic generation of decision reports, intelligent output of decision solutions, intelligent execution, and collection of feedback data by calling the training and inference results of large AI models.

2. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 1, characterized in that, The construction of the enterprise-wide resource platform described in step S1 specifically includes: A unified semantic mapping and ontology mapping engine enables semantic alignment and ontology association across data sources and system resources; data virtual mapping and logical orchestration technologies are used to standardize and encapsulate various heterogeneous resources into reusable and scalable core resource units, which are then incorporated into the unified management of the enterprise's full-domain resource platform, forming a standardized, ontological, and service-oriented core resource set; The multimodal data assets include data warehouses, data lakes, business system databases, enterprise-related documents, images, audio, and video. The expert knowledge base includes the enterprise's private domain knowledge base, business knowledge logic, enterprise operating indicators, and historical decision-making cases; The business systems include MES system, OMS system, CRM system, OA system, ERP system, and BI system; The unified semantic mapping and ontology mapping engine constructs an enterprise-level unified data dictionary and ontology vocabulary based on the inherent attributes of resources and business characteristics.

3. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 1, characterized in that, The construction of an ontology-driven enterprise global dynamic knowledge graph in step S2 specifically includes: After multiple processing operations such as entity disambiguation, cross-source alignment, knowledge fusion and ontology mapping, an enterprise global knowledge graph containing real-time dynamic knowledge is constructed and stored in a graph database to realize the structured organization, dynamic updating and efficient retrieval of knowledge. The core modeling system of the enterprise global dynamic knowledge graph includes an entity type system, a relationship type system, an entity attribute system, and a business action strategy system. The entity type system includes core enterprise entities, business process entities, product entities, customer entities, and hierarchical entity structures. The relationship type system includes one-to-one, one-to-many, and many-to-many association patterns and semantic constraints; The entity attribute system includes basic attributes, business attributes, and status attributes; The business action strategy system includes association rules, algorithm programs, and system API interfaces; Entity extraction and relation extraction are implemented using a deep learning model based on an attention mechanism.

4. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 1, characterized in that, Step S3, which involves building an AI large-scale model training and inference engine, specifically includes: The large model three-stage training engine employs a specifically designed large model three-stage training algorithm, which includes: Narrowing the knowledge graph retrieval scope: Based on the predefined entity type and relation type matching rules in the enterprise global dynamic knowledge graph, knowledge graph subgraphs related to the target reasoning task are selected. Specifically, this includes obtaining the text description of the target reasoning task, parsing the core entities and relationships of the task, comparing them with the entity type and relation type matching rules of the knowledge graph, extracting knowledge graph subgraphs containing the core entities and relationships based on efficient graph database retrieval algorithms, and automatically filtering irrelevant nodes and invalid relation links; finally, embedding the knowledge graph subgraphs into the corresponding graph representation using a graph neural network. Screening high-quality paths in the knowledge graph: For the reasoning paths in the subgraphs of the knowledge graph, a dual-dimensional evaluation of relevance and completeness is conducted. The relevance evaluation is quantitatively represented by calculating the semantic association weight between the reasoning path and the target task, while the completeness evaluation is completed by determining the degree to which the reasoning path covers the core elements of the target task. By fusing the relevance and completeness indicators of the paths, a path evaluation function is constructed using a neural network model. Based on the evaluation results, high-quality knowledge paths that meet the requirements of the target reasoning task are ranked and screened to reduce interference from redundant paths. The large model is trained and optimized in two steps: First, supervised fine-tuning is performed on the basic AI model using sample data corresponding to the high-quality knowledge paths. The sample data includes the input question, knowledge graph path, and standard output results. By minimizing the objective function, the model's adaptability to knowledge graph information is improved. Second, the preference alignment algorithm is used to optimize the model after SFT fine-tuning. By using manually labeled preference sample sets and optimizing the model's reward function, the model output is made to better meet the needs of enterprise decision-making.

5. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 4, characterized in that, The node representation update logic of the graph neural network is as follows: at the first... Layers, nodes The representation is updated as follows: ； in, Represents a node In the Layer representation; The relation matrix representing the learning process; Indicates aggregation operation; Indicates an update operation; initial representation and Random initialization, The union of the neighborhoods of all subject entities. Let r be the set of subgraph triples; r represents the graph relationship between node e' and node e.

6. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 5, characterized in that, The path evaluation function is defined as follows: ； MLP stands for Multilayer Perceptron; Indicates the activation function; This represents the input query text; Indicates the cumulative value of the path; Estimate the value of the path; Among them, the cumulative value of the path The calculation method is as follows: ； in, Indicates from arrive The path set; Represents a node Update logic, , Represents a node The total number of floors; Path Prediction Value The calculation method is as follows: ； in, It is a feedforward neural network. This indicates concatenating two vectors; The path evaluation function is trained using the cross-entropy loss function, which is defined as follows: ； in, Indicates query The correct set of entities; This indicates the set of other entities in the query subgraph that are incorrect.

7. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 6, characterized in that, The objective function for the supervised fine-tuning is: ； in, These represent the model parameters of a basic AI model. Represented as The key reasoning path for screening; The parameter is AI large models in known and Time generation The conditional probability; This represents the training dataset; This represents the expected value calculation for the training samples; The objective function for preference alignment is: ； in, This indicates a preference for aligning datasets; Indicates the most important preferred path; This represents a less important secondary path randomly sampled in the subgraph; This represents the temperature parameter, used to control the degree to which the model deviates from the reference model; This represents the sigmoid function.

8. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 7, characterized in that, Step S3 also includes: The hybrid reasoning engine adopts a combination of knowledge graph symbolic reasoning and large-model semantic reasoning. The high-performance computing relies on GPU clusters to support parallel acceleration of model training and inference.

9. The enterprise intelligent decision-making method based on large models and knowledge graphs according to claim 7, characterized in that, Step S3, which involves building an intelligent application based on a Multi-Agent collaborative architecture, specifically includes: The Hybrid Agent architecture is adopted, and the execution logic is as follows: Complex tasks are broken down and executed by B-Agent based on the ReAct framework, while simple tasks are quickly responded to by S-Agent based on the Workflow pattern. By leveraging feedback learning agents to build a closed-loop feedback iteration mechanism of demand analysis, task execution, and reflection iteration, the performance of large AI models can be continuously optimized, supporting enterprises in real-time insight, intelligent decision-making, and business empowerment.

10. An enterprise intelligent decision-making system based on large models and knowledge graphs, characterized in that, An enterprise intelligent decision-making method based on large models and knowledge graphs as described in any one of claims 1 to 9, comprising: The enterprise-wide resource platform construction module is used to integrate and logically fuse multimodal data assets, expert knowledge bases, and various data, knowledge, and system resources corresponding to enterprise business systems. It achieves semantic alignment and ontology association across data sources and system resources through a unified semantic mapping and ontology mapping engine. It adopts data virtual mapping and logical orchestration technology to standardize and encapsulate various heterogeneous resources into reusable and scalable core resource units, which are then incorporated into the unified management of the enterprise-wide resource platform to form a standardized, ontological, and service-oriented core resource set. The Enterprise Global Dynamic Knowledge Graph Construction Module is used to build an ontology-based enterprise global dynamic knowledge graph. Specifically, it includes: an element extraction unit, which extracts core elements needed for enterprise decision-making through multi-dimensional extraction of entities, relationships, attributes, and business actions; a fusion and alignment unit, which performs entity disambiguation and cross-source alignment on the extracted core element information; and a graph construction and storage unit, which, after knowledge fusion and ontology mapping, constructs an enterprise global knowledge graph containing real-time dynamic knowledge and stores it in a graph database, achieving structured organization, dynamic updates, and efficient retrieval of knowledge. The AI ​​large-scale model training and inference framework module is used to build the training and inference framework and complete the training and inference of the AI ​​large-scale model. Specifically, it includes: a three-stage training engine unit, which is configured with three-stage training algorithm logic, including logic for narrowing the search scope, logic for selecting high-quality paths, and logic for two-step training optimization; a hybrid inference engine unit, which adopts inference logic that combines knowledge graph symbolic inference and large-scale model semantic inference; a high-performance computing unit, which relies on GPU clusters to provide parallel acceleration support; and a model fine-tuning unit, which is used to realize supervised fine-tuning and preference alignment optimization of the large-scale model. The multi-Agent collaborative intelligent application module is configured with an intelligent data analysis agent, an intelligent report generation agent, a decision inference and execution agent, and a feedback learning agent. Each agent is implemented based on a hybrid agent architecture that integrates the ReAct framework and the Workflow mode. The intelligent application module is used to call the inference results of the AI ​​large model training inference module to realize intelligent analysis of enterprise business data, automatic generation of decision reports, intelligent output of decision solutions, intelligent execution, and collection of feedback data.

Images