Large model business logic processing method and system based on workflow engine and domain knowledge fusion

Abstract

The invention relates to a large model business logic processing method and system based on workflow engine and domain knowledge fusion, and is suitable for automatic processing of complex multi-node and multi-branch business processes. According to the method, natural language input of a user is analyzed through a large language model, a service intention is recognized, and the service intention is converted into an executable task process through a workflow engine. And the workflow engine dynamically adjusts an execution path according to rules and data in the domain knowledge graph to realize efficient parallel processing of tasks. The system integrates a workflow engine, a domain knowledge graph and a large language model, supports real-time data processing, rule matching, conflict resolution and decision optimization, is suitable for the fields of water conservancy, medical treatment, finance and the like, automatically generates and executes a complex business process, and optimizes a decision process. The system improves the accuracy and execution efficiency of business decision through deep fusion of domain knowledge.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and business process management, and in particular to a method and system for processing large-scale business logic based on the fusion of workflow engine and domain knowledge. Background Technology

[0002] With the continuous development of artificial intelligence technology, especially the rise of large language models (LLMs), intelligent applications based on large models have gradually become powerful tools for solving complex business problems in many fields. However, although large language models perform well in natural language processing and knowledge mining, they still have many limitations when dealing with complex business logic in certain professional fields. In particular, traditional models and technical solutions have not been able to fully meet practical needs in areas such as the management of multi-node, multi-branch business processes, the matching and conflict resolution of business rules, and the integration of complex domain knowledge.

[0003] I. Shortcomings of existing technologies

[0004] 1. Limitations of large-scale language models

[0005] Currently, while large-scale language models can perform efficient natural language processing and generate fluent text output, they still face several key challenges when handling complex business logic with domain-specific knowledge. First, the knowledge base of large-scale language models primarily comes from massive amounts of text data, lacking a deep understanding of specific rules and business processes within a particular domain. For example, in water conservancy business processes, complex reservoir scheduling rules, ecological protection, and flood control scheduling are involved, and existing large-scale language models struggle to accurately identify and handle the priority and conflicts of these domain-specific rules. Second, although large-scale language models are strong in understanding textual intent, they lack the ability to effectively manage cross-system and multi-node tasks. For instance, business processes in industries such as water conservancy, healthcare, and finance often involve data flow and task collaboration between multiple systems, and existing large-scale language models struggle to effectively manage these task nodes and data dependencies.

[0006] 2. Limitations of traditional workflow systems

[0007] Traditional workflow systems typically rely on pre-set rules and process templates, exhibiting strong structure and standardization, making them suitable for handling fixed and simple tasks. However, they fall short when facing complex and ever-changing business scenarios. First, traditional workflow systems lack natural language understanding capabilities, hindering their ability to flexibly adapt to extreme conditions and complex business scenarios. For example, in water conservancy scheduling, traditional workflow systems cannot quickly identify and automatically adjust business processes in the face of unconventional situations such as once-in-a-century floods; timely responses to such extreme conditions place higher demands on the intelligence capabilities of the business system. Second, traditional workflow systems usually rely on manually preset rules and lack real-time learning and knowledge updates. When business needs change (such as adjustments to regulations and policies, or changes in weather conditions), traditional systems often require manual updates to the rule base, resulting in long update cycles and a high risk of errors. Finally, traditional workflow systems have weak support for cross-system collaboration and struggle to seamlessly integrate with external data sources or specialized computing models, limiting the automation and intelligence of business processes.

[0008] 3. Fragmentation of professional knowledge

[0009] Currently, knowledge in many industry sectors (such as water conservancy, healthcare, and finance) is scattered across different data sources, including structured data (such as real-time monitoring data and business records in databases), semi-structured data (such as industry standards and scheduling plans), and unstructured data (such as expert experience and historical cases). This knowledge often exists in fragmented form, lacking a unified knowledge management and retrieval mechanism. This not only leads to low knowledge reuse rates but also makes it difficult to effectively integrate various types of knowledge when dealing with complex business logic. For example, in the water conservancy field, real-time water level data, historical scheduling data, and expert opinions are often stored and processed independently, lacking effective cross-data source collaboration and knowledge integration, resulting in insufficiently comprehensive and accurate basis for business decisions.

[0010] 4. Issues related to dynamic changes in business rules

[0011] With environmental changes and the continuous emergence of new technologies, business rules in many industries require frequent adjustments and updates. For example, in the water conservancy sector, water resource allocation rules may need to be adjusted in real time based on actual conditions due to changes in infrastructure such as reservoirs and rivers, as well as climate conditions. Traditional business systems often rely on manual intervention for rule revision and updates, which is not only inefficient but also prone to errors. Currently, large-scale language models lack sensitivity to rule updates and cannot effectively integrate new rule knowledge into the system in a timely manner, resulting in a slow response time to changes.

[0012] II. Application Scenarios of Existing Technologies

[0013] Existing technologies are primarily applied to simple business processes and data processing tasks. For example, traditional workflow systems can automate simple business processes such as order processing and approval workflows, but they still have many limitations for complex business scenarios (such as water conservancy scheduling and medical diagnosis). Especially in fields such as water conservancy, healthcare, and finance, these business processes typically involve multiple complex conditional judgments, rule matching, and cross-system data collaboration. Existing technologies struggle to effectively address issues such as business logic processing, rule conflicts, and task scheduling in these complex tasks.

[0014] III. Technical Problem to be Solved by the Invention

[0015] The core technical challenge of this invention is how to overcome the shortcomings of existing technologies, such as insufficient intelligence, difficulty in handling rule conflicts, and low efficiency in multi-node business process collaboration, through deep integration of workflow engines and domain knowledge graphs, thereby improving the capabilities of large language models in complex business processing within specialized domains. Specifically, this invention aims to solve the following key problems:

[0016] 1. Flexibility and dynamic adaptability of process management: How to achieve flexible management of multi-branch and multi-node business processes through a workflow engine, and dynamically adjust the workflow under extreme conditions to ensure the smooth execution of complex business processes.

[0017] 2. Efficient identification and handling of rule priorities and conflicts: How to dynamically manage the priority of rules in complex business scenarios through domain knowledge graphs and rule engines, ensure coordination and accurate matching of priorities among multiple rules, and avoid errors caused by rule conflicts.

[0018] 3. Deep integration of multi-source data and professional knowledge: How to integrate structured data, semi-structured data and unstructured knowledge by establishing a unified knowledge access interface to form an efficient knowledge graph that supports the workflow engine in providing intelligent support for business decisions.

[0019] 4. System Collaboration and Dynamic Adaptation: How to overcome the limitations of traditional workflow systems that lack collaboration with external systems, and improve the overall collaborative capabilities of the system by seamlessly integrating various professional computing models and external data sources, and realize the dynamic updating and automatic learning of business rules.

[0020] The technical solution provided by this invention breaks through the bottleneck of traditional technologies in intelligent business processing by deeply integrating workflow engines, domain knowledge graphs and large-scale language models, and provides efficient and intelligent solutions for complex business processes in the industry, especially in professional fields such as water conservancy, medical care and finance, where it has important application value. Summary of the Invention

[0021] This invention proposes a method and system for processing large-scale business logic based on the fusion of workflow engines and domain knowledge. The core idea of ​​this method is to achieve efficient and automated processing of complex business processes by deeply integrating large-scale language models (LLMs) with workflow engines and domain knowledge graphs. In traditional business processing systems, handling complex business logic often faces the problem of coordinating multiple nodes and branches. Furthermore, the lack of flexible dynamic rule adjustment mechanisms and real-time invocation of professional knowledge greatly limits the efficiency and accuracy of the system.

[0022] To address these issues, this invention processes business logic through the following key steps: First, a large-scale language model is used to parse the natural language input by the user and identify the business intent within it. Next, a workflow engine transforms the identified business intent into a specific execution flow, with the system dynamically adjusting the execution path based on the rules and data of the domain knowledge graph. During execution, the workflow engine effectively improves task execution efficiency through parallel task processing and dependency management between nodes. Finally, the system integrates the execution results of each node with the large-scale language model to generate the final business response.

[0023] The innovation of this approach lies in its use of domain knowledge graphs to enable the workflow engine to flexibly adjust execution paths when faced with complex business rules and task nodes, avoiding the rigid performance of traditional workflow systems under extreme conditions. Simultaneously, the system can dynamically handle multiple conflicting business rules and optimize decisions based on real-time data and historical experience, further improving the accuracy and responsiveness of business logic processing.

[0024] In one possible implementation, this invention provides a large-model business logic processing method based on the fusion of workflow engine and domain knowledge. The core idea of ​​this method is to parse user natural language input using a large language model (LLM), identify the business intent within it, and transform these intents into an executable workflow process, which is then executed by the workflow engine. Based on preset business rules and dependencies between task nodes, the workflow engine sequentially completes tasks such as accessing the data system, applying business rules, and invoking the domain model through dynamic path generation and parallel execution. Finally, the large model integrates the execution results of each node to generate a natural language response that meets the requirements of the business domain.

[0025] In one possible implementation, during the user input business intent recognition stage, the large language model first extracts business entities (such as "reservoir," "flood warning," etc.) and operational intents (such as "start scheduling," "execute allocation," etc.) from the input text using natural language processing techniques (such as named entity recognition, intent classification, etc.). This information is used to determine the business scenario requested by the user and helps identify the type of business process. Simultaneously, the large language model determines the priority of business rules based on the domain knowledge graph, ensuring that various rules are executed according to priority, thereby avoiding conflicts and errors. For example, there may be a conflict between flood control scheduling and irrigation needs; through the priority determination mechanism, flood control scheduling rules will be executed first.

[0026] In another possible implementation, the present invention dynamically generates executable business processes through a workflow engine. The workflow engine maps identified business intents to preset workflow templates and generates a complete executable process. This process may involve multiple task nodes (such as data acquisition, model calculation, rule determination, etc.), and the execution order and logical relationships of these nodes are dynamically adjusted by the workflow engine to adapt to the actual situation. During execution, the workflow engine automatically selects the execution path based on the status and results of the task nodes. For example, during emergency dispatch, if the reservoir's water level exceeds a set threshold, the system will automatically adjust to an emergency dispatch path, bypassing the conventional process, and formulate an emergency response plan through model calculation.

[0027] In one possible implementation, the workflow engine not only supports dynamic path generation but also enables the parallel execution of multiple independent node tasks. In some business processes, multiple task nodes can run simultaneously; for example, in water resource scheduling, multiple tasks such as reservoir water level monitoring and rainfall prediction can be executed concurrently, thereby improving processing efficiency. Furthermore, the workflow engine can handle complex data dependencies between nodes, ensuring that each task node executes correctly based on the results from upstream nodes.

[0028] In another possible implementation, the workflow engine further enhances its dynamic adaptability to business rules and knowledge by integrating with a domain knowledge graph. The domain knowledge graph not only encompasses structured information such as business rules and industry standards, but also includes unstructured knowledge such as historical cases and expert experience. This knowledge is integrated with the workflow engine through standardized interfaces, enabling the workflow engine to dynamically adjust execution paths based on real-time data and handle abnormal situations. For example, during emergency response, the workflow engine can quickly switch to an emergency handling branch based on information such as current water levels and rainfall, and generate an emergency plan based on expert experience recorded in the domain knowledge graph, ensuring the scientific validity and rationality of the results.

[0029] In one possible implementation, the process of mapping business intents to executable workflow processes also includes a conflict detection and resolution mechanism for multiple intent inputs. When user input contains multiple conflicting business intents, the system can automatically identify and resolve the conflicts according to preset rules. For example, when the input simultaneously contains the intents of "ensuring urban water supply" and "prioritizing ecological water replenishment," the system uses priority rules to determine that the need to ensure urban water supply will be processed first, ensuring that the system's execution results meet actual needs.

[0030] Furthermore, the present invention maps business intent into an executable process, which includes the following key steps:

[0031] 1. Foundation Layer: Establishes a mapping relationship between business intents and workflow templates based on a domain knowledge graph. This mapping relationship combines multi-dimensional information such as business entities, rule constraints, and operation processes to ensure that business intents can be accurately mapped to preset workflow templates.

[0032] 2. Conflict Detection Layer: Detects potential conflicts between multiple intents using a fuzzy matching algorithm. The algorithm includes:

[0033] Calculate the cosine similarity of intent vectors in the semantic space to quantify the semantic similarity between different business intents;

[0034] Assess the degree of overlap in rule constraints and determine whether there are conflicts between rules with different intentions;

[0035] Quantify the resource consumption conflict index to assess the extent to which multiple business intentions may consume system resources (such as computing, storage, network bandwidth, etc.).

[0036] 3. Priority Processing Layer: Dynamically prioritizes detected conflicts. The prioritization criteria include:

[0037] The preset rule weights in the domain knowledge graph determine the priority of different business rules based on domain expertise;

[0038] The urgency index of real-time business data assesses the urgency of tasks based on the current business environment and data status.

[0039] Historical conflict resolution records, referenced from historical decision-making records, can be used to optimize conflict resolution strategies.

[0040] 4. Execution Layer: Automatically generates workflow definitions that include conflict resolution parameters. These parameters include:

[0041] Branch isolation flags for parallel execution are used to ensure that parallel execution of multiple business processes does not cause resource conflicts;

[0042] The mutex lock mechanism for critical resources ensures that critical tasks are processed first when multiple tasks share limited resources;

[0043] The exception handling callback function configuration enables fault tolerance processing based on the preset callback mechanism when the system encounters an exception, ensuring system stability and business continuity.

[0044] Through the above technical solutions, this invention possesses high flexibility and intelligence in handling complex business logic. It can automatically adjust the execution flow based on the user's input business intent and improve the accuracy and efficiency of business logic processing through the integration of domain knowledge. Especially when dealing with complex processes with multiple nodes and branches, the workflow engine of this invention can effectively manage process nodes, optimize execution paths, and reduce error rates through intelligent decision-making mechanisms, achieving efficient and accurate business processing.

[0045] Based on the above technical solutions, the present invention significantly improves the accuracy and flexibility of large language models in processing complex business logic by deeply integrating workflow engines and domain knowledge. Specifically, the present invention solves the problems of difficulty in identifying business rule priorities and conflicts, difficulty in generating dynamic paths for complex processes, and inefficiency in processing multi-source data that exist in current technologies.

[0046] First, by using the workflow engine's dynamic path generation algorithm, combined with rules from the domain knowledge graph and real-time business data, the optimal path can be intelligently selected, and the execution order of task nodes can be automatically adjusted according to the actual situation, ensuring that the entire process can run efficiently in complex environments.

[0047] Secondly, through a business rule priority management mechanism, this invention can perform version control and priority sorting of business rules, adjust rule priorities in real time during business processes, avoid rule conflicts, and improve the accuracy of rule matching.

[0048] Finally, the domain knowledge fusion module of this invention can efficiently integrate structured data and unstructured knowledge, enabling unified access and rapid retrieval of multi-source data. This not only improves the model's computational power but also enhances the transparency and interpretability of the decision-making process. By combining historical cases and expert experience, business rules can be quickly adjusted according to new business needs and environmental changes.

[0049] Through the above technical solutions, this invention can overcome the limitations of traditional large models in handling complex business in professional fields, improve the automation level and intelligent decision-making capabilities of business processes, and is especially suitable for multi-node and multi-branch collaborative business processing in professional fields such as water conservancy, medical care, and finance, thus promoting the development of intelligent business processing in various industries.

[0050] This invention solves several key problems in complex business logic processing by combining workflow engines, domain knowledge graphs, and large-scale language models, and mainly achieves the following breakthrough technical advantages:

[0051] 1. Improve processing accuracy

[0052] By integrating workflow engines with domain knowledge, the "illusion" phenomenon and rule matching errors that often occur when large models handle complex business logic are effectively resolved. In actual water conservancy business scenarios, testing and comparison showed that the accuracy of business logic processing increased from 65% to 92% compared to a pure large model solution. This significantly reduced the workload of manual review, improved business processing efficiency, and ensured the accuracy and efficiency of business execution.

[0053] 2. Implement dynamic rule management

[0054] This invention provides a visual management interface for business rules, supporting online editing, priority adjustment, and version control. When policies, regulations, or business needs change, the system can quickly update relevant rules, reducing the rule update time from an average of two weeks to one to two days. This improvement ensures that the system always meets the latest business requirements, greatly enhancing the flexibility and responsiveness of rule management.

[0055] 3. Supports multi-system collaboration

[0056] This invention achieves seamless collaboration between large models, workflow engines, and external systems (such as water conservancy monitoring systems and specialized computing models) through a unified knowledge access interface and mechanism model invocation adaptation layer. For example, in the scenario of joint scheduling of reservoir groups, the system can simultaneously call real-time data from multiple reservoirs and different hydrodynamic models, improving the comprehensiveness and accuracy of business processing and ensuring data integration and decision optimization.

[0057] 4. Enhance process explainability

[0058] The workflow engine tracks and records the execution status of each process node in real time, and generates detailed decision explanations based on a large model, making the entire business process transparent. Users can query the basis and data source for each decision through the system. This not only improves the transparency of decision-making, but also meets the regulatory requirements for traceability and compliance of business processes in fields such as water conservancy and healthcare.

[0059] 5. Lower the implementation threshold

[0060] This invention provides graphical process design tools and domain-specific model templates, enabling non-technical personnel to quickly configure business processes and define rules. In grassroots applications, complex business systems can be built and run without the need for extensive technical support, significantly reducing implementation costs and technical barriers in specialized fields. This makes the invention widely applicable to various industries, lowering the threshold for intelligent transformation. Attached Figure Description

[0061] Figure 1 The system architecture diagram shows the five-layer architecture of the system and the interaction relationships between each layer;

[0062] Figure 2 The intent recognition and workflow matching flowchart illustrates the intent recognition and mapping process from user input to workflow execution.

[0063] Figure 3 A diagram illustrating the execution of multi-node business logic demonstrates how to manage and execute multi-node tasks in parallel within complex business processes.

[0064] Figure 4 The domain knowledge fusion mechanism diagram describes how structured and unstructured data are fused to support decision-making through knowledge graphs.

[0065] Figure 5 The flowchart for exception handling and result feedback illustrates how exceptions are handled during process execution and how the results are ultimately integrated and output. Detailed Implementation

[0066] To better understand the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0067] In a specific implementation, this invention provides a method and system for automating complex business logic processing through deep integration of a workflow engine and a domain knowledge graph, combined with a large language model (LLM). This technical solution enables the system to handle complex business processes with multiple branches and nodes, and optimizes the decision-making process with the support of real-time data. In specific implementation, the processing steps of this invention unfold step-by-step according to user requirement parsing, workflow generation and execution, knowledge graph-assisted decision-making, and result integration and feedback, ensuring the efficiency and flexibility of business processing.

[0068] First, the system begins with the user's natural language input, parsing it using a large language model. Through techniques such as named entity recognition and intent classification, it extracts key business entities and operational intents. At this stage, by integrating with a domain knowledge graph, the system further identifies the priority of business rules and identifies potential rule conflicts, ensuring that the system can automatically resolve conflicts based on priority and avoid errors during processing.

[0069] Next, after intent recognition, the system maps the identified business intent to the executable flow of the workflow engine. This process is not a static mapping, but rather a dynamic generation of workflows that meet actual needs based on real-time business data and changes in the external environment. The workflow engine's execution module, through parallel processing and dynamic path selection, can efficiently schedule tasks among nodes, making full use of computing resources and ensuring the efficient execution of business processes.

[0070] Throughout the workflow execution process, the system leverages rules and historical experience from the domain knowledge graph to optimize and manage dependencies between multiple nodes, ensuring that each execution node makes decisions based on the latest data and rules. The key role of the domain knowledge graph in this invention lies in providing the workflow engine with real-time business rule updates, conflict resolution, and decision support, ensuring the system can react quickly to changes in the external environment.

[0071] Finally, the results from each node's execution are aggregated and integrated. A large-scale language model transforms the data output from each node into a natural language response that meets the user's needs. This response not only includes the conclusions of business decisions but also provides the basis and data sources for each decision, ensuring the transparency and traceability of the decision-making process.

[0072] The specific implementation of this invention effectively overcomes the limitations of traditional technologies in complex business processing through a highly integrated workflow engine, domain knowledge graph, and large-scale language model, providing a flexible, intelligent, and efficient solution.

[0073] This invention provides an efficient and intelligent method for handling complex multi-node, multi-branch business processes by deeply integrating workflow engines and domain knowledge graphs. The specific implementation is as follows:

[0074] 1. System Architecture and Working Principle

[0075] like Figure 1 As shown, the overall system architecture of this invention is divided into five main layers: user interaction layer, intent recognition layer, workflow engine layer, domain knowledge layer, and data support layer. The functions and workflows of each layer are described in detail below.

[0076] 1.1. User Interaction Layer

[0077] Users submit natural language input to the system through the interaction layer, such as "Initiate flood control scheduling of the reservoir group." This layer is mainly responsible for receiving and passing user input to the next-level intent recognition layer. The system interacts with users through a web interface or mobile application, providing a simple and intuitive operating interface.

[0078] 1.2. Intent Recognition Layer

[0079] At the intent recognition layer, the system uses a finely tuned Large Language Model (LLM) to parse user input. Through Named Entity Recognition (NER) and intent classification techniques, the system can extract key information from the input (such as "reservoir group" and "flood control scheduling"). Furthermore, the system can incorporate domain knowledge graphs to enhance the accuracy of intent recognition. For example, when "reservoir group" is recognized, the system will further determine the scheduling target and scope based on the reservoir entities and their attributes in the domain knowledge graph.

[0080] 1.3. Workflow Engine Layer

[0081] Once the business intent is identified, the system generates an executable workflow at the workflow engine layer. Through integration with a domain knowledge graph, the workflow engine can dynamically adjust the execution path. For example, in water conservancy scheduling, the system generates the optimal path based on real-time data such as current reservoir water level and rainfall. Figure 3 This demonstrates how a workflow engine efficiently executes tasks through multi-node parallel processing and dynamic path adjustment. In this process, the workflow engine automatically identifies data dependencies within the process and dynamically selects appropriate task nodes for execution.

[0082] 1.4. Domain Knowledge Layer

[0083] The domain knowledge layer provides support for the system through knowledge graphs. Knowledge graphs not only cover structured data (such as reservoir capacity and flow rate), but also unstructured data such as industry rules, historical cases, and expert experience. Figure 4 The demonstration showcased the working mechanism of the domain knowledge graph. It connects business rules, historical cases, and other knowledge with the workflow engine through standardized interfaces, enabling the system to make decisions based on real-time data and historical experience during execution. For example, in the event of a flood warning, the system automatically invokes flood prevention rules, selecting the appropriate processing path based on rule priorities provided by the knowledge graph and expert suggestions.

[0084] 1.5. Data Support Layer

[0085] The data support layer is responsible for storing and accessing structured and unstructured data. This layer includes real-time monitoring data (such as water levels, flow rates, and rainfall), business records, and historical case studies. Through standardized data access interfaces, the workflow engine can efficiently access data and perform further processing as needed. Figure 1 The data support layer ensures the system's computing and storage capabilities, supporting the integration and analysis of business data from multiple fields.

[0086] 2. Specific Implementation Steps

[0087] 2.1. User Input Parsing and Intent Recognition

[0088] Users input natural language commands through the interaction layer, such as: "Initiate flood control scheduling and conduct joint scheduling of the reservoir group." This command first passes through the intent recognition layer. The system uses a finely tuned large-scale language model to parse the input text, extracting key business entities (such as "flood control scheduling" and "joint scheduling of the reservoir group") and operational intent. The system further confirms the task's objectives and scope using information such as reservoirs and scheduling rules from the domain knowledge graph.

[0089] 2.2. Intent Mapping to Workflow Engine

[0090] After intent recognition, the system maps the business intent to a preset workflow template in the workflow engine. Taking the water conservancy field as an example, if the intent is "start flood control scheduling", the system will map it to a series of task nodes: "monitor water level - start flood evolution model - generate scheduling plan - issue instructions". Figure 2 This demonstrates the mapping process, where the workflow engine automatically generates an executable flow based on the recognition results and configures relevant parameters for each node.

[0091] 2.3 Workflow Execution and Dynamic Path Selection

[0092] The workflow engine executes tasks based on preset templates and dynamically selects execution paths based on real-time data. For example, in water conservancy scheduling, when the reservoir water level exceeds a set threshold, the system will automatically switch to the emergency scheduling branch. Figure 3 This demonstrates how the workflow engine handles multi-node tasks and dynamic path generation. Through parallel execution and data dependency management, the workflow engine can process a large number of tasks in a short time, improving business processing efficiency.

[0093] 2.4. Domain Knowledge Support and Decision Optimization

[0094] During task execution, the workflow engine invokes business rules and expert experience from the domain knowledge graph to support decision-making. For example, in flood control scheduling, the system optimizes the scheduling plan based on information such as reservoir capacity, flow scheduling rules, and historical cases from the knowledge graph, ensuring the rationality of scheduling decisions. Figure 4 It demonstrates how to use domain knowledge graphs combined with real-time data to make intelligent decisions and adjust business processes.

[0095] 2.5. Exception Handling and Result Feedback

[0096] If an anomaly occurs at a task node during execution (such as missing water level data or calculation errors), the system will trigger the anomaly handling mechanism. Figure 5 The system demonstrates an exception handling process. When an exception is detected, the system executes retry, rollback, or repair strategies to ensure the stability and reliability of the process. Finally, the execution results of all nodes are collected and used to generate a natural language response through a large language model, providing users with the necessary business decision-making information.

[0097] 3. Results Output and Optimization

[0098] The system will output results in multimodal formats (text, charts, reports, etc.) according to user needs. In the scenario of water conservancy scheduling, the report generated by the system may include information such as reservoir scheduling plans, flood discharge volume, and execution time. Users can choose the presentation method and details of the report, and the system can also automatically optimize the report content according to user needs to ensure that it clearly and accurately expresses business decisions and execution processes.

[0099] This invention integrates a workflow engine with a domain knowledge graph and combines it with a large-scale language model to automate and intelligently process complex business processes. Whether in water conservancy, healthcare, finance, or other industries, the system can efficiently handle multi-node, multi-branch business processes, dynamically adapt to changes in business rules, and ensure system reliability through exception handling and feedback mechanisms. Through this invention, business process execution becomes more intelligent, flexible, and efficient, suitable for complex scenarios in various professional fields.

[0100] Example 1:

[0101] The embodiments of this invention are based on a distributed computing environment. By combining a workflow engine with a domain knowledge graph, it utilizes a large language model (LLM) to efficiently process complex multi-node, multi-branch business processes. In this process, through high-level hardware and software integration, this invention supports the automated processing of complex business logic in specialized domains and improves the system's decision-making accuracy and processing efficiency. The specific implementation steps of this invention are as follows.

[0102] The system implementation of this invention is based on a distributed computing environment to support complex business logic processing requirements. The specific hardware and software configuration requirements are as follows:

[0103] 1. Hardware Configuration Requirements

[0104] Computational Nodes: To meet the computational needs of large model inference, workflow engine, and knowledge graph, the system's computational nodes are configured as follows:

[0105] Quantity: At least 3 distributed computing nodes, supporting horizontal scaling to 10+ nodes, dynamically adjusted according to business complexity.

[0106] Processors: Each node is configured with at least two Intel Xeon Gold 6348 (28 cores) or equivalent CPUs that support multi-threaded parallel computing.

[0107] Memory: Each node is configured with at least 256GB of DDR4 ECC memory to meet the needs of large model inference, workflow engine and knowledge graph loading.

[0108] Acceleration chips: Each node is configured with at least one NVIDIA A100 (80GB VRAM) or equivalent GPU for accelerated computation of intent recognition, result integration and mechanism models of large models.

[0109] Storage nodes: Adopting HDFS distributed storage architecture, with a total storage capacity of at least 10TB, supporting high-throughput read and write, and equipped with an SSD caching layer (single node ≥ 1TB NVMe SSD) for high-frequency access to business rule base, process templates and intermediate result caching.

[0110] Network environment: To ensure efficient distributed computing and low-latency node communication, the system adopts 100Gbps InfiniBand high-speed interconnection, and the communication latency between nodes is kept within 1ms; it provides Ethernet interfaces of 10Gbps or higher to support high-concurrency access requests from user terminals.

[0111] Monitoring nodes: The system independently deploys monitoring servers (CPU ≥ 8 cores, memory ≥ 32GB), running distributed monitoring tools (such as Prometheus + Grafana) to collect the CPU, memory, GPU utilization and task execution status of each node in real time, ensuring the stable operation of the system.

[0112] 2. Software Dependencies

[0113] The implementation of this invention relies on the following software tools and frameworks to support large model inference, workflow engine scheduling, multi-source knowledge fusion, and distributed execution management:

[0114] The underlying operating system: Both compute nodes and control nodes use Ubuntu 22.04LTS, supporting containerized deployment and cross-node resource scheduling. The control node is configured with a Kubernetes 1.24+ container orchestration platform to manage distributed resources.

[0115] Large model service framework: It adopts vLLM 0.3.0+ or ​​TensorRT-LLM 0.6.0+ inference engine, supports large model fine-tuning and intent recognition optimization, encapsulates large model interface based on microservice architecture, and supports load balancing and dynamic scaling.

[0116] Workflow Engine and Rule Engine: The system's core workflow engine adopts the integrated difyAI workflow engine, and the rule engine adopts Drools 8.44+, which supports priority configuration and visual management of business rules (integrated rule editor).

[0117] Knowledge storage and access components: Neo4j 5.10+ graph database is used to store the relationship graph of domain entities and business rules; relational database (such as MySQL 8.0+) is used to store structured business data and process execution logs; multi-source knowledge access middleware provides a unified interface for databases, file systems and third-party knowledge services to the system through RESTful API.

[0118] Distributed execution and monitoring tools: Distributed task scheduling uses Apache DolphinScheduler 3.2.0+, and task dependency management is implemented through this tool; the monitoring and alerting system uses Prometheus 2.45+ and Grafana 10.2+, which supports real-time monitoring of system status; log management uses the ELK Stack (Elasticsearch 8.8++, Logstash 8.8++, Kibana 8.8+), which supports full-process log traceability and problem localization.

[0119] Development and runtime environment: The core modules for large model interaction and algorithm implementation use Python 3.9, while the workflow engine and rule engine extensions use Java 17+; containerization tools Docker 24.0+ and Kubernetes 1.26+ ensure consistent service orchestration and environment management; the interface protocols uniformly adopt HTTP / 2 and gRPC protocols to ensure efficient cross-module and cross-system communication.

[0120] During implementation, the workflow engine, domain knowledge graph, and large model of this invention will work together to ensure efficient processing of complex business logic. The following are the specific steps executed by the system:

[0121] S101 System Initialization Phase

[0122] First, the system models domain-specific business processes using visual modeling tools (such as Dify), and builds template libraries for fields such as water conservancy, healthcare, and finance. Templates support parameterized configuration, such as "early warning level" in the water conservancy template and "symptom type" in the healthcare template, ensuring that templates can be dynamically configured according to actual business needs. Through standardized workflow templates, the system can adapt to changing business requirements and automatically instantiate workflows based on the scenario.

[0123] At this stage, a business rule base needs to be built through a combination of manual sorting and intelligent extraction. Rules such as "When the reservoir water level reaches the flood control limit and the rainfall in the next 24 hours exceeds 50mm, flood control scheduling will be prioritized" will be encoded using the Drools rule engine and marked with rule priorities (e.g., flood control > irrigation > ecology). The knowledge graph construction integrates domain entities (e.g., reservoirs, rivers, monitoring stations), relationships (e.g., upstream and downstream relationships, affiliation relationships), and attributes (e.g., reservoir capacity, water level). Through Natural Language Processing (NLP) technology, unstructured knowledge such as historical cases and expert experience is integrated into the graph to establish a "rule-case-model" relationship.

[0124] Meanwhile, specialized computational models (such as reservoir group joint scheduling models and hydrodynamic models) will be encapsulated as microservices using Docker container technology and registered with the service center. This allows the system to call external models for computation as needed and seamlessly integrate them with business processes.

[0125] S102 Intent Processing Stage

[0126] Users input natural language commands (e.g., "Initiate flood warning and dispatch for the XX river basin") via a web interface or mobile device. The system transmits the user commands to the large model service node through a reverse proxy server (such as Nginx). The fine-tuned large language model (such as a BERT-based water conservancy model) performs semantic parsing on the input commands, identifies key information such as "XX river basin," "flood warning," and "initiate dispatch," and extracts business entities and intent elements.

[0127] Using Named Entity Recognition (NER) technology, the system extracts business entities (such as "XX Reservoir" and "XX Hydrological Station") and related operational intents (such as "flood control scheduling") from the parsing results. Combined with the domain knowledge graph, the system will also query the attributes of the entities (such as "flood control limit water level of the reservoir") and supplement the contextual information of the intent to construct a complete intent description framework.

[0128] The next step after intent processing is workflow task sequence generation. Through cosine similarity calculation and rule-based reasoning, the system matches the extracted intent elements with templates in the workflow template library, selects a suitable workflow template, and generates an execution task sequence. For example, in a flood control scheduling scenario, the system will automatically generate a task sequence such as "water level monitoring—flood model calculation—scheduling plan generation".

[0129] S103 Process Execution Phase

[0130] Once the intent is parsed and mapped to the workflow engine, the system instantiates the workflow based on the workflow template. The workflow engine allocates corresponding computing resources to each node according to the task sequence. For example, the system might allocate two GPU cores to the flood control scheduling process instance for flood evolution model calculations and one CPU core to the rule judgment. During this process, the workflow engine ensures that each task node executes sequentially and dynamically adjusts the execution path based on real-time data or business rules.

[0131] The workflow engine uses an event-driven mechanism (such as Apache Kafka) to coordinate data flow between nodes. When a node (such as the "water level monitoring" node) completes its task, it automatically triggers the execution of subsequent nodes (such as the "flood model calculation" node). The system can also call services across systems, such as calling a pre-registered flood forecasting model service in the "flood evolution calculation" node to obtain and process relevant data.

[0132] S104 Knowledge Application Stage

[0133] During execution, the workflow engine assesses each decision node based on the business rule base. For example, at the "whether to initiate flood discharge" node, the system will determine whether flood control scheduling needs to be initiated based on the current water level and rainfall. If multiple rules match the same condition, the system resolves conflicts based on priority and historical cases in the knowledge graph, thereby generating a reasonable decision solution.

[0134] The workflow engine also calls specialized computational models through a mechanism model adaptation layer. For example, in a water resource allocation scenario, it calls the SWAT model to perform watershed water balance calculations to further optimize decisions. Based on model results and rule judgments, the workflow engine dynamically adjusts the execution path. For instance, when the predicted peak flood flow exceeds the design standard, the system automatically adjusts the execution path and enters the emergency scheduling branch.

[0135] S105 Result Processing Stage

[0136] Once all task nodes have completed their execution, the system will aggregate the outputs from each node and perform semantic fusion using a large model, transforming structured data into natural language descriptions. For example, it might generate a report stating: "It is recommended that gates 1 and 3 of the XX Reservoir be opened, and the flood discharge flow controlled at 200 m³ / s."3 At the same time, the system will also verify the reasonableness of the output results to ensure that they comply with business rules and industry standards.

[0137] All data, rule matching records, and model call logs during execution are stored in Elasticsearch, creating a traceable execution archive. Users can query the basis and data source for each decision. Finally, the system provides feedback to users in a multimodal manner, including visual charts, report summaries, and decision instructions.

[0138] This implementation details each stage of the system, covering hardware configuration, software environment, system initialization, intent processing, workflow execution, knowledge application, and result output. By deeply integrating domain knowledge graphs with large-scale models, this invention effectively enhances the intelligence level of business logic processing and is widely applicable to the automated management and intelligent decision-making of complex business processes in industries such as water conservancy, healthcare, and finance.

[0139] Example 2:

[0140] To further illustrate the specific applications of this invention, the following example, using the flood control and dispatching scenario of the Chaohu Lake basin reservoir group, details the practical application of this invention in complex business logic processing. In this scenario, the system receives natural language input commands, parses business intents, generates workflows, and dynamically executes various dispatching tasks, ultimately achieving automated and efficient flood warning and reservoir dispatching.

[0141] In this embodiment, the user input the following command through the system's natural language interface: "Currently, there is widespread heavy rainfall in the Chaohu Lake basin, with an hourly rainfall of 80mm. The water level of Dongpu Reservoir is 28.3m (flood control limit level 28m), the water level upstream of Chaohu Sluice Gate is 12.5m, and the water level of Yuxi River downstream of the sluice gate is 10.93m; the water level at Datong Station on the Yangtze River is 8.2m (lower than the Chaohu Lake water level). Please initiate the joint scheduling of the reservoir group within the basin and prioritize the discharge of floodwaters." After receiving the command, the system immediately began processing the input information and executing business logic based on a large-scale language model and domain knowledge graph.

[0142] First, the system uses a finely tuned large model, combined with a knowledge graph of the Chaohu Lake basin, to parse the natural language input by the user. Through semantic analysis, the system extracts key information from the instructions, including the water level of Dongpu Reservoir, basin rainfall, Yangtze River water level, and Yuxi River flow. Furthermore, the system, based on relevant clauses of the "Chaohu Sluice Gate Control and Operation Measures," determines that the current water level and flow meet the conditions for initiating joint scheduling. The core business intent identified by the system is to initiate the joint scheduling of the basin's reservoir group, prioritizing flood discharge.

[0143] After extracting the intent elements, the system further detected no conflicts between all intents, ensuring the consistency and rationality of the scheduling tasks. Next, the system selected the "Joint Scheduling of Reservoirs Exceeding Flood Limits" template from the preset template library using an intent-process mapping algorithm, and generated a specific task sequence. The task sequence includes steps such as initializing scheduling parameters, real-time data acquisition from multiple reservoirs, flood evolution simulation, scheduling rule matching, multi-scheme generation, scheme optimization, cross-sluice collaborative execution, and execution monitoring.

[0144] During workflow execution, the workflow engine creates an instance and allocates corresponding computing resources. Specifically, two GPU cores are allocated for flood control model computation, and one CPU core is used for rule judgment, ensuring efficient execution of each task node. Simultaneously, the system connects to the real-time databases of three reservoirs in the Chaohu Lake basin to provide accurate data support for subsequent tasks.

[0145] During actual node execution, the system first collects real-time data from multiple reservoirs through a unified interface, including the water level of Dongpu Reservoir (28.3m, exceeding the flood limit by 0.3m), the water level upstream of Chaohu Sluice Gate (12.5m), and the water level downstream of the gate (10.93m). In addition, the system also acquires relevant data such as the water level at the Datong Station of the Yangtze River and the downstream flow of the Yuxi River, providing necessary basic data for subsequent scheduling decisions. Based on this data, the system enters the flood evolution simulation stage. In this stage, the workflow engine calls the flood forecast model through the adaptation layer, inputting basin rainfall forecast data and the current water level information of each reservoir to simulate the flood evolution results. The model output indicates that the water level of Dongpu Reservoir is expected to rise to 28.42m within the next 6 hours, and the downstream river flow will exceed the warning value.

[0146] In the subsequent scheduling rule matching phase, the system determines whether to initiate flood discharge operations based on the priority rules in Article 3.1.3 of the "Chaohu Sluice Control and Operation Method". According to the system's calculation results and preset rules, the water level at Chaohu Sluice (12.5m) is higher than the water level at the downstream sluice (10.93m), and the water level at the Datong Station of the Yangtze River is lower, meeting the condition of "floodwaters arriving first and being able to drain by themselves". Therefore, the system decides to initiate flood discharge operations at Chaohu Sluice, maximizing the flood discharge flow according to the principle of "symmetrical opening from the middle to both sides".

[0147] Next, the system combined the rules and model results to generate three scheduling schemes as shown in Table 1. Each scheme corresponds to different reservoir discharge rates, gate opening methods, and total flood discharge rates. Schemes A, B, and C respectively configured the discharge rate of Dongpu Reservoir and the opening methods of Chaohu Gate and Yuxi Gate to ensure that each scheme meets the prescribed scheduling requirements and satisfies the principle of prioritizing flood control safety.

[0148] Table 1: Three scheduling schemes were generated by combining the rules and model results.

[0149]

[0150] After the plan was generated, the system selected the optimal scheduling plan based on compliance and effect predictions and began execution. Ultimately, the scheduling instructions included opening two gates at Dongpu Reservoir, releasing a flow rate of 100 m³ / h. 3 / s; The Chaohu Sluice Gate opened all 16 gates in a symmetrical manner from the middle to both sides, with a discharge capacity of 1370 m³ / s. 3 / s; Yuxi Sluice Gate 14 gates fully open, discharge capacity 1170m³ / s 3 / s. The system predicts that in 6 hours, the water level of Dongpu Reservoir will drop to 28.0m, returning to the flood control limit; the water level upstream of Chaohu Sluice Gate will drop to 11.8m; and the downstream flow will stabilize at 380m³. 3 / s, below the warning value of 400m 3 / s.

[0151] Through this embodiment, the present invention fully utilizes the synergistic effect of large model reasoning, knowledge graph and workflow engine in flood control scheduling in Chaohu Lake Basin to realize the automated execution and intelligent decision-making of scheduling tasks, thereby effectively improving flood response capabilities, reducing the risk of manual intervention, and improving system response speed and decision accuracy.

[0152] This embodiment demonstrates how the present invention, in the field of water conservancy, particularly in flood control scheduling scenarios, handles the complex business logic of joint scheduling of multiple reservoirs by combining an intelligent workflow engine with a domain knowledge graph. The system automatically generates and executes scheduling plans through multiple steps, including natural language parsing, intent recognition, rule matching, and flood simulation, achieving high efficiency, scientific rigor, and real-time performance in scheduling decisions. This not only reduces manual intervention and improves work efficiency but also enhances the ability to respond to sudden floods, demonstrating the broad application potential of the present invention in practical business operations.

[0153] Example 3: Specific Implementation of Multi-Intent Conflict Detection and Automatic Scheduling Priority Allocation Method

[0154] This embodiment, based on the technical solution proposed in this invention, combines a workflow engine, a domain knowledge graph, and a large-scale language model to explain in detail how to achieve intelligent decision-making and automated scheduling of complex business processes through conflict detection, priority sorting, and parameterized process generation mechanisms when user input contains multiple business intentions.

[0155] In a typical water conservancy business scenario, the user simultaneously inputs the following commands through the interactive interface:

[0156] "The current regional rainstorm warning is in effect. Please activate the reservoir flood control scheduling while also ensuring ecological water replenishment."

[0157] After receiving the natural language instruction, the system performs the following processing steps:

[0158] S301: Multi-Intent Resolution and Mapping

[0159] The system uses a large language model to perform semantic analysis on user input text, identifying two main business intents: "flood control scheduling" and "ecological water replenishment." It also extracts business entities (e.g., "reservoir"), operation types (e.g., "scheduling," "water replenishment"), and contextual factors (e.g., "rainstorm warning"). Subsequently, by combining workflow templates defined in the domain knowledge graph, the two intents are mapped to standard workflow template nodes, and resource dependencies are recorded.

[0160] S302: Conflict Detection Mechanism

[0161] In the conflict detection phase, the system introduces a semantic vector model to vectorize business intent and determines whether a conflict exists based on the following three types of indicators:

[0162] Semantic similarity metric: Calculate the cosine similarity between "flood control scheduling" and "ecological water replenishment" in the semantic space. If the similarity is higher than a set threshold (e.g., 0.85), it indicates that the two are highly correlated and may have overlapping targets or resources.

[0163] Rule constraint overlap: Analyze the business rules applicable to two workflows in the domain knowledge graph. If contradictions are found in the target water level range, gate opening and closing strategies, etc., it is determined to be a rule conflict.

[0164] Resource Conflict Index: Simulates the occupation of key resources (such as reservoir scheduling capacity and number of gate channels) to assess whether there are conflict bottlenecks when two tasks are performed simultaneously.

[0165] If any of the above indicators exceeds the set threshold, the system will determine that there is a potential conflict in the multi-intent input and it needs to proceed to the next priority processing stage.

[0166] S303: Dynamic Priority Ranking Mechanism

[0167] The system prioritizes and ranks two conflicting intentions based on the following three information sources:

[0168] Rule weight value: This refers to the preset priority weights of flood control rules and ecological water replenishment rules in the domain knowledge graph. Based on policy documents and water conservancy industry standards, flood control rules have a higher priority and are assigned a higher initial weight.

[0169] Real-time urgency indicators: By combining current meteorological and hydrological monitoring data (such as rainfall, reservoir water levels, and flow velocity), the urgency of flood control tasks is dynamically assessed. For example, when a red rainstorm warning is issued, the system will raise the urgency factor of flood control scheduling to the highest level.

[0170] Historical conflict resolution records: The system reviews similar historical scheduling cases recorded in the domain knowledge graph (such as scheduling behavior in similar areas in the past 5 years). If flood prevention is selected as the priority in most cases, the priority score will be further increased.

[0171] Ultimately, the system calculates the scheduling priority of each intention using a weighted scoring method, and the ranking result is: flood control scheduling > ecological water replenishment.

[0172] S304: Conflict Resolution and Process Definition Generation

[0173] Based on the sorting results, the system automatically generates a parameterized workflow definition, including the following:

[0174] Process isolation parameters: Generate independent execution branches for two business processes and set parallel flags to ensure that high-priority tasks (flood control) are executed first, while ecological water replenishment tasks can be set to delayed start or conditional trigger.

[0175] Key resource mutual exclusion lock mechanism: Mutual exclusion locks are allocated to shared resources (such as control gates) to ensure that such resources are not occupied by ecological water replenishment processes before the flood control process is completed.

[0176] Abnormal callback configuration: If an abnormality occurs during flood control execution (such as equipment failure, sudden water level change, etc.), the system will automatically trigger a preset callback function to pause or interrupt the ecological water replenishment process and call the emergency water replenishment strategy.

[0177] S305: Process Deployment and Execution

[0178] The system automatically deploys the generated workflow to the execution engine and schedules each node step by step according to priority. During scheduling, the system continuously monitors key indicators (such as water level changes, task completion status, etc.) and automatically adjusts the task path when necessary. The final execution result will be processed into a natural language report through a large language model and fed back to the user, including execution actions, conflict handling instructions, key parameters, and prediction results.

[0179] The mechanism described in this embodiment is not only applicable to handling flood control and water replenishment conflicts in water conservancy scenarios, but can also be extended to fields such as medical resource scheduling (e.g., conflicts between surgery and examination scheduling), energy allocation, emergency logistics, and urban traffic management. The system can adaptively adjust scheduling priorities and process generation strategies based on knowledge graphs and rule settings in different domains, achieving general intelligent handling and execution control capabilities for multi-intent conflicts.

[0180] This invention has been described in detail through specific embodiments; however, it should be understood that these embodiments are merely examples and not intended to limit the scope of the invention. Any changes, substitutions, and modifications made to this invention by those skilled in the art without departing from the inventive concept are within the scope of protection of this invention. The scope of protection of this invention should be defined by the appended claims.

[0181] Therefore, the technical solutions provided by this invention are not limited to the specific steps and device configurations described in the above embodiments. Any changes, modifications, and equivalent substitutions that conform to the essential characteristics of this invention should be included within the scope of protection of this invention. In particular, the large-scale language model business logic processing method and system based on the workflow engine and domain knowledge graph of this invention has highly innovative and practical technical features, steps, and processes, and can be effectively applied to complex business process management and decision support in multiple fields. All changes or modifications to the principles, structures, and methods within the technical solutions of this invention should be considered within the scope of patent protection.

Claims

1. A method for processing large-scale business logic based on the fusion of workflow engine and domain knowledge, characterized in that, Includes the following steps: S1. Identify the business intent in the user's natural language input through a large-scale language model; S2. Map the identified business intents to executable processes in the workflow engine; S3, the workflow engine executes according to process nodes: accessing data systems, applying business rules, and calling domain models; S4. Collect the execution results of each node and integrate them with the large model to generate the final response.

2. The method according to claim 1, characterized in that, The business intent identification mentioned in step S1 includes: Extract business entities and operational intentions; Determine the priority of business rules; The rules base is used to identify and resolve conflicts between intents.

3. The method according to claim 1, characterized in that, The workflow engine execution in step S2 includes: Dynamically select execution paths and support parallel node tasks; Handle data dependencies between nodes; The management node executes the status and handles exceptions.

4. The method according to claim 1, characterized in that, The step S2, which maps business intents to executable processes, includes: Establish a mapping between business intents and workflow templates based on domain knowledge graphs; Perform conflict detection and prioritization on multi-intent inputs; Automatically generate workflow definitions that include parameter configurations.

5. The method according to claim 1, characterized in that, When the workflow engine executes node tasks in step S3, it includes: A unified data access interface is used to obtain structured and unstructured data. Business rules are matched and applied based on a rules engine; Call upon external domain computing models for data processing and decision support.

6. The method according to claim 1, characterized in that, Step S4, which involves collecting execution results and integrating them to generate the final response, includes: Cleaning and standardizing multi-source heterogeneous data; Utilizing large-scale language models for result interpretation and knowledge fusion; Generate natural language responses that conform to the business domain.

7. The method according to claim 4, characterized in that, The domain knowledge graph includes: Business entities and their attribute relationships; Business rules and constraints; Historical cases, expert experience, and knowledge from other fields.

8. The method according to claim 5, characterized in that, The workflow engine supports the following: Real-time monitoring of the execution status of process nodes; Exception handling mechanisms, including automatic retries and rollbacks; Record execution trajectory for auditing and tracing purposes.

9. The method according to claim 6, characterized in that, The large language model is fine-tuned using domain knowledge, including: supervised fine-tuning using domain-specific texts; Construct a domain terminology dictionary and entity recognition model; A prompt template for optimizing the model for specific business scenarios.

10. A large-scale business logic processing system based on the fusion of workflow engine and domain knowledge, characterized in that, include: The intent recognition module is used to identify the business intent in the user's natural language input; The process mapping module is used to map business intents into executable workflow processes; The workflow execution module executes tasks according to process nodes and handles dependencies between nodes. The results integration module is used to collect the execution results of each node and generate the final response.

Images