A complex information system and data platform application fusion method and system based on a large language model agent and a computer readable storage medium

Abstract

The application discloses a complex information system and data middle platform application fusion method and system based on a large language model intelligent agent and a computer readable storage medium. The method is that an intelligent core performs semantic analysis and event element extraction on a complex task in a natural language form issued by a complex information system, and converts the complex task into structured task data. Macroscopic node arrangement is performed based on a preset decision chain template, and a double-graph knowledge enhancement mechanism is used in each node to generate a sub-task sequence. Candidate tools are recalled from a tool set to trigger optimal tool execution. The optimal tool execution results corresponding to each node are received, which are comprehensively arranged with external knowledge to generate a structured decision result and return to the complex information system. Task whole-process data is backflowed and stored in a multi-modal database for iterative optimization. The problems of application form difference between the complex information system and the data middle platform application, static solidification calling, insufficient task decomposition reliability and low tool matching efficiency are solved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and information systems, specifically relating to a method and system for integrating complex information systems and data middleware applications based on large language model intelligent agents, as well as a computer-readable storage medium. Background Technology

[0002] With the rapid development of information technology, two types of large-scale systems are commonly found across various industries, built independently: one is a complex information system centered on "process-driven" principles; the other is a data platform centered on "service-driven" principles. Data platforms aggregate and integrate heterogeneous data from multiple sources, providing discrete and independent data products, AI models, business models, and data services such as modular data. The full realization of their functionality depends on the invocation and collaboration of business systems. Therefore, how to enable data platforms to better serve process-driven application systems has become a critical issue that urgently needs to be addressed.

[0003] Functionally, a data platform can provide agile data support for complex information systems, enhancing their intelligence level. However, in actual deployments, the two have long been relatively separate, making deep application integration difficult. Existing solutions for their integration suffer from the following problems: 1. The structural differences in application formats lead to difficulties in service embedding. The structural differences in application formats make it difficult for data services from the data platform to be accurately and dynamically embedded into the business processes of complex information systems.

[0004] 2. The problem of insufficient reliability of large language models in decomposing complex tasks directly. When complex tasks are directly input into large language models for decomposition, the following defects exist: (1) poor process compliance, the sub-task sequence generated by the large language model is easy to deviate from the standard business process; (2) domain semantic deviation, the general large language model lacks vertical domain knowledge, and the generated results are prone to "illusion"; (3) lack of logical constraints, the large language model has difficulty in autonomously capturing the temporal dependencies and mutual exclusion constraints between sub-tasks, resulting in disordered execution chain logic; (4) imbalance between flexibility and controllability, the decomposition process driven purely by the large language model lacks human intervention interface, and the controllability is poor.

[0005] 3. The problem of low matching efficiency and poor accuracy when the data platform tools are large in scale. The data platform usually registers hundreds to thousands of tools. The existing solutions have the following problems: (1) Context window limitation, large language models cannot process the full description of tools at one time; (2) Insufficient accuracy of single matching strategy, relying only on keyword matching or semantic matching, it is impossible to capture the multi-dimensional correspondence between sub-tasks and tools; (3) Unstable sorting results, lacking a differentiable multi-criteria sorting mechanism.

[0006] 4. Existing solutions pose significant challenges to existing systems, including high intrusiveness and stability risks. Complex information systems and data platforms are both massive undertakings; deep modifications to either system introduce stability risks and increase construction costs. Existing integration solutions often require substantial alterations to existing systems, making implementation difficult while ensuring their stability.

[0007] In summary, there is an urgent need for an application integration method and system that can dynamically and intelligently connect complex information systems and data platforms, and possesses good reliability, controllability, and scalability. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, the present invention aims to provide a method and system for integrating complex information systems and data platform applications based on large language model intelligent agents, as well as a computer-readable storage medium, to solve the problems of application form differences, static and fixed calls, insufficient reliability of task decomposition, and low efficiency of tool matching between existing complex information systems and data platform applications.

[0009] This invention is achieved through the following technical solution: A method for integrating complex information systems based on large language model intelligent agents with data platform applications includes: Step 1: An intelligent hub is deployed between the complex information system and the data platform application. The intent understanding module in the intelligent hub performs semantic parsing and event element extraction on the complex tasks in natural language form issued by the complex information system, and converts the complex tasks in natural language form into structured task data. Step 2: The task decomposition module in the intelligent hub receives the structured task data in Step 1, performs macro-node arrangement based on the preset decision chain template, and generates sub-task sequences within each node using the dual-graph knowledge enhancement mechanism. Step 3: The tool recommendation and invocation module in the intelligent hub receives the sub-task sequence in step 2, recalls candidate tools from the tool set of the data platform application, sorts the candidate tools in a differentiable manner, and triggers the execution of the optimal tool in a function call manner. Step 4: The integrated service module in the intelligent hub receives the execution results of the optimal tools corresponding to each node, integrates the execution results of the optimal tools corresponding to each node with external knowledge, generates structured decision results, and returns them to the complex information system. Step 5: The feedback optimization module in the intelligent hub will feed back the data of the entire task process and store it in the multimodal database of the data platform application to support the iterative optimization of the intelligent hub.

[0010] Furthermore, in step 1, the intent understanding module processing steps are as follows: Step 1.1: Use the large language model in the semantic parsing unit to perform semantic parsing on the complex task in the form of natural language, and extract event elements and constraints. The event elements include at least time, location, task type, and event level. Step 1.2: Use the dialogue state tracking unit to generate a state vector based on the multi-turn dialogue history encoding. When event elements are missing, ask the user follow-up questions through guided dialogue. Step 1.3: Using the structured output unit, organize the extracted elements into structured task data according to the predefined structured output format.

[0011] Furthermore, in step 2, the task decomposition module processes the following steps: Step 2.1: Retrieve candidate decision chains from a pre-set decision chain template library based on semantic vector matching using the decision chain matching unit; Step 2.2: The node sequence defined by the selected decision chain is used as the macro framework for task execution by the macro node orchestration unit. The node sequence includes linear serial processing stages. Step 2.3: By calling the dual-graph knowledge enhancement mechanism in each node through the micro-node decomposition unit, the macro-task of the current node is decomposed into a sequence of specific sub-tasks that can be directly executed by the tool.

[0012] Furthermore, in step 2.3, the processing steps of the dual-graph knowledge enhancement mechanism are as follows: Step 2.3.1: Call the large language model and combine it with semantic similarity calculation to extract the set of key entities from the task text of the current node; Step 2.3.2: Retrieve temporal dependency and logical constraint triples from the constraint relationship graph, retrieve semantic association triples from the domain knowledge graph, and aggregate them into structured knowledge fragments by D-hop adjacency subgraph; Step 2.3.3: Retrieve text fragments from the unstructured knowledge base of the data platform application using retrieval enhancement generation technology, and then splice and merge the structured knowledge fragments with the unstructured knowledge fragments to generate a knowledge-enhanced document; Step 2.3.4: Input the knowledge-enhanced document as prompt words into the large language model, guide the large language model to generate a thought chain corresponding to the logical structure of the decision chain nodes, and output a specific sub-task sequence.

[0013] Furthermore, in step 3, the tool recommendation and invocation module processing steps are as follows: Step 3.1: Using a semantic encoder, the semantic recall unit encodes the subtask text and the metadata description of each tool in the toolset, sorts them in descending order of cosine similarity, and recalls the Top-K candidate tools from the toolset. Step 3.2: Construct a multidimensional feature vector for the candidate tools through a differentiable sorting unit and input it into the LapSum differentiable sorting operator. Output a comprehensive sorting value and select the optimal tool in descending order. Step 3.3: The large language model automatically fills in the parameters according to the input specifications of the selected optimal tool and the subtask context through the request generation unit, and generates a call request that conforms to the predefined JSON pattern; Step 3.4: Parse the call request through the interface triggering unit, trigger the execution of the corresponding microservice through the unified API gateway of the toolset, and return the standardized execution result; For each subtask, repeat sub-steps 3.1 to 3.4 until all subtasks in the current decision chain are assigned and the corresponding optimal tool is executed, thus obtaining the set of optimal tool execution results for each node.

[0014] Furthermore, in step 4, the processing steps of the integrated service module are as follows: Step 4.1: Retrieve supplementary information related to the final output of the current task from the knowledge base, vector library, and graph library of the data platform application through the knowledge retrieval unit; Step 4.2: The result fusion unit uses a large language model to integrate the output results of each node tool with the supplementary information obtained from the retrieval, generates structured decision results according to a predefined result template, and returns them to the complex information system through a standardized interface.

[0015] Furthermore, the feedback optimization module feeds back the entire task process data and stores it in the multimodal database. The processing steps are as follows: Step 5.1: The process data and result data of the entire task are fed back through the feedback channel via the data feedback unit; Step 5.2: Perform ETL cleaning on the returned data through the data processing unit, convert the cleaned data into a vectorized representation and add classification labels to generate training samples in a standard format; Step 5.3: Store the training samples in the standard format into the vector library, knowledge base, and graph library of the multimodal database, respectively; Step 5.4: Based on the accumulated training samples in the vector library, knowledge base, and graph library, the model iteration unit periodically triggers the fine-tuning training of the large language model, and at the same time updates the embedding representation of the dual graph in the task decomposition module and the differentiable sortable feature weights in the tool recommendation and invocation module, so as to realize the optimization and iteration of the strategies of each module of the intelligent hub.

[0016] Furthermore, prior to step 1, the process includes building a toolset within the data platform application, with the following steps: The metadata of various tools within the data platform application is uniformly registered to the tool registration center. The metadata includes at least the tool's name, functional semantic description, tags, input parameter specifications, and output format specifications. Based on the input parameter specifications and output format specifications stored in the tool registry, the operation interface of each type of tool is encapsulated into a structured API interface through the interface encapsulation unit, providing a unified calling specification to shield the heterogeneity of each tool. During the process of the intelligent hub calling various tools to execute tasks, the continuous learning unit collects logs and performance data generated by the tools and feeds the logs and performance data back to the tool training and optimization process to achieve continuous improvement of tool capabilities.

[0017] A complex information system and data platform application fusion system based on a large language model intelligent agent includes a complex information system, an intelligent hub, and a data platform application; the intelligent hub is deployed between the complex information system and the data platform application and includes an intent understanding module, a task decomposition module, a tool recommendation and invocation module, a comprehensive service module, and a feedback optimization module; The intent understanding module is used to perform semantic parsing and event element extraction on complex tasks in natural language form issued by the complex information system, and to convert the complex tasks in natural language form into structured task data. The task decomposition module is used to receive structured task data obtained by the intent understanding module, perform macro-node arrangement based on a preset decision chain template, and generate sub-task sequences within each node using a dual-graph knowledge enhancement mechanism. The tool recommendation and invocation module is used to receive the sub-task sequence obtained by the task decomposition module, recall candidate tools from the tool set of the data platform application, sort the candidate tools, and trigger the execution of the optimal tool in the form of a function call. The integrated service module is used to receive the optimal tool execution results corresponding to each node obtained by the tool recommendation and invocation module, integrate and organize the optimal tool execution results corresponding to each node with external knowledge, generate structured decision results, and return them to the complex information system. The feedback optimization module is used to feed back the data of the entire task process and store it in the multimodal database of the data platform application to support the iterative optimization of the intelligent hub.

[0018] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the steps of any of the fusion methods described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, through the design of an independently deployed intelligent hub, dynamically and intelligently connects the "linear" business processes of complex information systems with the "point-like" data services of data platform applications, effectively solving the problem of service embedding difficulties caused by the structural differences in the application forms of the two. By introducing a two-layer task decomposition architecture based on decision chains (i.e., macro-level node orchestration units and micro-level node decomposition units) and a dual-graph knowledge enhancement mechanism (constraint relationship graph and domain knowledge graph), it effectively solves the reliability and standardization problems of directly decomposing tasks using large language models. By constructing a toolset and adopting a two-stage alignment mechanism (semantic recall unit and differentiable sorting unit), it effectively solves the matching accuracy and efficiency problems of large-scale tool calls. Through a modular intelligent hub intermediate layer design, it effectively solves the problem of high costs associated with modifying existing systems in current integration solutions. The method and system of this invention can be applied to various fields such as government affairs, dispatching, emergency response, healthcare, and finance that require embedding data platform data services into complex business processes. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall system framework of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the intended module structure of the present invention.

[0022] Figure 3 This is a schematic diagram of the task decomposition module structure of the present invention.

[0023] Figure 4 This is a schematic diagram of the recommended and calling module structure for the tools of this invention.

[0024] Figure 5 This is a schematic diagram of the integrated service module structure of the present invention.

[0025] Figure 6 This is a schematic diagram of the feedback optimization module structure of the present invention.

[0026] Figure 7 This is a schematic diagram of the toolset structure in a data middle platform application. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention.

[0028] refer to Figure 1This paper presents a complex information system and data platform application fusion system based on a large language model intelligent agent. The system employs a three-layer architecture, comprising the complex information system side, the intelligent hub, and the data platform application side. The intelligent hub is independently deployed between the complex information system side and the data platform application side, serving as a bridge for dynamic integration at the application level. The intelligent hub receives complex tasks in natural language form from the complex information system and invokes tools from the data platform application toolkit to complete the tasks. Internally, it is organized according to a full-link logic of "perception—planning—invocation—execution—feedback," including an intent understanding module, a task decomposition module, a tool recommendation and invocation module, a comprehensive service module, and a feedback optimization module.

[0029] 1. Intelligent Central Hub The intelligent hub is the core component of this invention, serving as a crucial intermediate layer connecting complex information systems with data platform applications and enabling dynamic integration of their application layers. It employs a modular design, consisting of five interconnected functional modules forming an end-to-end processing pipeline. The output of one module serves as the input for the next, creating a complete processing chain.

[0030] 1.1 Intent Understanding Module The intent understanding module serves as the input point, receiving complex tasks in natural language form from the complex information system and converting them into structured task data. The intent understanding module comprises a semantic parsing unit, a dialogue state tracking unit, and a structured output unit, which operate sequentially: semantic parsing unit → dialogue state tracking unit → structured output unit.

[0031] (1) Semantic parsing unit The semantic parsing unit receives natural language task instructions from a complex information system, uses a large language model to perform semantic parsing, and extracts the event elements and constraints of the task. Event elements include at least attribute information such as time, location, task type, and event level. The semantic parsing unit guides the large language model to extract the corresponding content one by one according to predefined attribute slots through predefined prompt templates, ensuring the completeness and consistency of the extraction results.

[0032] (2) Dialogue state tracking unit The dialogue state tracking unit is used to solve the context association problem in multi-turn dialogue scenarios. When the event elements extracted by the semantic parsing unit are incomplete, the dialogue state tracking unit, based on the Transformer architecture dialogue state tracking model, asks the user for the missing elements through guided dialogue; at the same time, it encodes the multi-turn dialogue history into a state vector to ensure that the model can accurately associate the context in subsequent turns and avoid ambiguous references or semantic drift.

[0033] (3) Structured output unit The structured output unit organizes the element information processed by the semantic parsing unit and the dialogue state tracking unit into structured task data according to a predefined structured output format, which serves as the unified input to the task decomposition module.

[0034] 1.2 Task Decomposition Module The task decomposition module is the core module for achieving standardized and controllable decomposition of complex tasks. Its innovation lies in introducing an intelligent agent workflow engine and adopting a two-layer decomposition architecture that combines macro-level node orchestration with micro-level node decomposition. The task decomposition module consists of three units: a decision chain matching unit, a macro-level node orchestration unit, and a micro-level node decomposition unit, which are connected sequentially.

[0035] (1) Decision chain matching unit The decision chain matching unit retrieves the most suitable decision chain for the current task from a pre-built decision chain template library. This library stores multiple decision chain templates, each corresponding to a standard business process, predefined by domain experts based on relevant business specifications. After the intent understanding module outputs structured task data, the decision chain matching unit uses semantic vector matching technology, employing a Sentence-BERT semantic encoder to vectorize the task description and decision chain description. It calculates the semantic similarity between the structured task data and the description vectors of each candidate decision chain, retrieving the Top-N candidate decision chains for manual selection.

[0036] (2) Macro-node arrangement unit The macro-node orchestration unit uses the node sequence defined by the selected decision chain as the macro-skeleton for task execution. Each node corresponds to a standard processing stage; for example, for a general business scenario, the node sequence can be set as "situation assessment → solution formulation → plan generation → quality verification." Nodes form a linear serial structure, with the output of the previous node serving as the input of the next, constructing a continuous and controllable processing link. After each node completes, the system introduces a manual modification and confirmation interface to achieve closed-loop control with "human intervention in the loop."

[0037] (3) Micro-node decomposition unit The micro-node decomposition unit performs subtask decomposition within each macro-node, breaking down the macro-task of the current node into a sequence of specific subtasks that can be directly executed by the tool.

[0038] 1.3 Tool Recommendation and Invocation Module The tool recommendation and invocation module is used to accurately select the most suitable execution tool from the toolset on the data platform side for each subtask output by the task decomposition module, and trigger the actual execution of the selected tool. The tool recommendation and invocation module includes a semantic recall unit, a differentiable sorting unit, a request generation unit, and an interface triggering unit. The first two units complete the tool recommendation, and the latter two units complete the tool invocation. The four units are connected in sequence to form a complete tool alignment and invocation chain of "recall-sorting-request generation-execution trigger".

[0039] (1) Semantic Recall Unit The semantic recall unit is used to quickly filter out a small number of candidate tools that are most relevant to the semantics of the subtask from a toolset of hundreds to thousands.

[0040] (2) Differentiable sorting unit Differentiable sorting units are used to perform secondary fine sorting on the candidate toolset obtained in the recall phase, enabling multi-dimensional comprehensive decision-making.

[0041] (3) Request generation unit The request generation unit generates a call request conforming to a predefined JSON pattern based on the input specifications of the selected tool, using the large language model. In practice, the toolset pre-registers standardized input parameter descriptions for each tool, including parameter names, types, and value ranges. Before the call, the large language model automatically fills in the corresponding parameters based on the context of the subtask, generating a structured call request object.

[0042] (4) Interface triggering unit The interface triggering unit receives the JSON call request object output by the request generation unit, parses its content, and triggers the execution of the corresponding microservice through the unified API gateway of the toolset. After the microservice completes execution, it returns the structured result to the interface triggering unit, which then feeds the result back into the large language model as input for subsequent inference.

[0043] 1.4 Integrated Service Module The integrated service module is used to integrate and output the intermediate results from each node of the decision chain, including a knowledge retrieval unit and a result fusion unit.

[0044] (1) Knowledge retrieval unit: retrieves supplementary information and resources related to the final output of this task from the knowledge base, vector library and graph library of the data platform.

[0045] (2) Result Fusion Unit: The large language model is used to integrate the output results of each node tool with the supplementary information obtained from the retrieval, and a structured decision result is generated according to the predefined result template. The structured decision result is encapsulated according to the interface format agreed upon by the complex information system side, and returned to the complex information system side for presentation through a standardized interface.

[0046] 1.5 Feedback Optimization Module The feedback optimization module is used to realize the closed-loop learning and continuous evolution of the system, including a data feedback unit, a data processing unit, and a model iteration unit.

[0047] (1) Data feedback unit: After the complex information system completes the task processing by calling the decision results, it summarizes and evaluates the entire task process, and the process data and result data are fed back to the feedback optimization module through the feedback channel.

[0048] (2) Data processing unit: The returned data is processed sequentially by ETL, vectorization, labeling and other steps, and converted into a standard format that can be used for model training. The data is then stored in the multimodal databases such as the vector library, knowledge base, and graph library of the data platform.

[0049] (3) Model Iteration Unit: Based on the processed data, the fine-tuning of the large language model and the optimization of the strategies of each module of the intelligent hub are triggered periodically, so that the capabilities of the intelligent hub continue to grow with the use process.

[0050] 2. Data Platform Side: Toolset In addition to the existing infrastructure such as databases, knowledge bases, vector libraries, and graph libraries, this invention adds a "toolset" component to the data platform side, serving as a unified entry point for calling data platform tools. This toolset is developed based on a microservice architecture using Spring Cloud and is deployed as a software entity on the data platform side. The toolset includes a tool registry center, an interface encapsulation unit, and a continuous learning unit.

[0051] (1) Tool Registration Center: The metadata such as names, functional semantic descriptions, tags, input parameter specifications, and output format specifications of various tools such as data products, AI mini-models, business mini-models, and blocks in the data platform are uniformly registered to the tool registration center, so as to facilitate retrieval and scheduling by various modules of the intelligent hub.

[0052] (2) Interface encapsulation unit: encapsulates the operation interface of each type of tool into a structured API interface, provides a unified calling specification to the outside world, and shields the heterogeneity of tools.

[0053] (3) Continuous learning unit: Collect logs and performance data generated during the execution of the tool and feed them back to the training and optimization process of the tool to achieve continuous improvement of the tool's capabilities.

[0054] A method for integrating complex information systems based on large language model intelligent agents with data platform applications includes: Step 1: An intelligent hub is deployed between the complex information system and the data platform application. The intent understanding module in the intelligent hub performs semantic parsing and event element extraction on the complex tasks in natural language form issued by the complex information system, and converts the complex tasks in natural language form into structured task data. The intent understanding module processes the following steps: Step 1.1: Use the large language model in the semantic parsing unit to perform semantic parsing on the complex task in the form of natural language, and extract event elements and constraints. The event elements include at least time, location, task type, and event level. Step 1.2: Use the dialogue state tracking unit to generate a state vector based on the multi-turn dialogue history encoding. When event elements are missing, ask the user follow-up questions through guided dialogue. Step 1.3: Using the structured output unit, organize the extracted elements into structured task data according to the predefined structured output format.

[0055] Step 2: The task decomposition module in the intelligent hub receives the structured task data in Step 1, performs macro-node arrangement based on the preset decision chain template, and generates sub-task sequences within each node using the dual-graph knowledge enhancement mechanism. The task decomposition module processes tasks in the following ways: Step 2.1: Retrieve candidate decision chains from a pre-set decision chain template library based on semantic vector matching using the decision chain matching unit; Step 2.2: The node sequence defined by the selected decision chain is used as the macro framework for task execution by the macro node orchestration unit. The node sequence includes linear serial processing stages. Step 2.3: By calling the dual-graph knowledge enhancement mechanism in each node through the micro-node decomposition unit, the macro-task of the current node is decomposed into a sequence of specific sub-tasks that can be directly executed by the tool.

[0056] The core of the micro-node decomposition unit is the dual-graph knowledge enhancement mechanism, and the specific processing flow is as follows: Step 2.3.1, Entity Extraction: The micro-node decomposition unit receives the task text Tcurrent of the current node, calls the large language model and combines semantic similarity calculation to extract the set of key entities from the task text: Extract(·) is the defined entity extraction function that maps the task text to a set of entities.

[0057] Step 2.3.2, Dual-Graph Retrieval: This invention pre-constructs two types of knowledge graphs: (a) Constraint Relationship Graph The constraint graph stores the temporal dependencies and logical constraints in the task execution process in the form of triples. For example, the triple (task A, precedes, task B) means that task A must be executed before task B. The constraint graph is used to provide logical dependency guidance for subsequent task decomposition, ensuring that the generated subtask sequence meets the execution requirements.

[0058] (b) Domain knowledge graph Key entities and their semantic relationships within a vertical domain are stored in the form of triples. The domain knowledge graph serves as semantic enhancement information for entities, compensating for the structural deficiencies in vertical domain knowledge in general large models.

[0059] For each entity Search for entities in the two types of maps that are no more than [distance not specified]. Jump Adjacency Subgraph By merging these fragments, we obtain the knowledge segment of the entity: The sum of knowledge fragments from all entities yields the original set of structured knowledge fragments. Step 2.3.3, Knowledge-Enhanced Document Generation: Simultaneously, the micro-node decomposition unit uses retrieval enhancement generation technology to retrieve text fragments (Krag) related to the current task from the unstructured knowledge base of the data platform. This structured knowledge is then used to generate the document. With unstructured knowledge The data is then combined and merged to generate the final knowledge-enhanced document. in This indicates an information concatenation operation. The merged document. It encompasses the execution logic dependencies in the constraint relationship graph, the semantic associations in the domain knowledge graph, and the background knowledge of RAG retrieval, which together constitute a complete task context.

[0060] Step 2.3.4, Subtask Generation: Micro-node decomposition units enhance knowledge documentation. As a prompt word The core component is the input to the large language model, which guides the large language model to generate a thought chain corresponding to the logical structure of the decision chain nodes, decomposing the current node task into a specific sequence of sub-tasks: in To integrate Structured cue words.

[0061] Step 3: The tool recommendation and invocation module in the intelligent hub receives the sub-task sequence in step 2, recalls candidate tools from the tool set of the data platform application, sorts the candidate tools in a differentiable manner, and triggers the execution of the optimal tool in a function call manner. The steps for tool recommendation and module invocation are as follows: Step 3.1: Using a semantic encoder, the semantic recall unit encodes the subtask text and the metadata description of each tool in the toolset, sorts them in descending order of cosine similarity, and recalls the Top-K candidate tools from the toolset. In practice, the Qwen3-Embedding-0.6B lightweight semantic encoder is used to process the subtask texts. and each tool in the toolset The metadata description is encoded by Meta to obtain a vector: Calculate the cosine similarity between the two as a semantic matching score: Sort by similarity score in descending order and quickly recall from the toolset. One candidate tool (preferred in this embodiment) ), forming a candidate set This stage ensures high recall, providing a high-quality candidate set for subsequent fine-tuning; Step 3.2: Construct a multidimensional feature vector for the candidate tools through a differentiable sorting unit and input it into the LapSum differentiable sorting operator. Output a comprehensive sorting value and select the optimal tool in descending order. The LapSum differentiable sorting algorithm is used, and the processing procedure for differentiable sortable units is as follows: ①Feature construction: Construct a multi-dimensional feature vector for each candidate tool. The feature dimensions should include at least semantic similarity, keyword matching degree, and entity association degree, and normalize all features to the [0,1] interval; ② Differentiable ranking: Input the feature matrix of the candidate tool into the LapSum operator, combine it with the preset feature weights, and output the comprehensive ranking value of each tool. This value reflects the superiority of the tool in multiple dimensions. ③Optimal output: Select the optimal tool in descending order of sorting value as the execution tool for the current subtask.

[0062] Through a two-stage alignment mechanism, the tool recommendation module achieves a balance between recall and precision, effectively solving the accuracy and efficiency issues of matching large-scale toolsets.

[0063] Step 3.3: The large language model automatically fills in the parameters according to the input specifications of the selected optimal tool and the subtask context through the request generation unit, and generates a call request that conforms to the predefined JSON pattern; Step 3.4: Parse the call request through the interface triggering unit, trigger the execution of the corresponding microservice through the unified API gateway of the toolset, and return the standardized execution result; For each subtask, repeat sub-steps 3.1 to 3.4 until all subtasks in the current decision chain are assigned and the corresponding optimal tool is executed, thus obtaining the set of optimal tool execution results for each node.

[0064] Step 4: The integrated service module in the intelligent hub receives the execution results of the optimal tools corresponding to each node, integrates the execution results of the optimal tools corresponding to each node with external knowledge, generates structured decision results, and returns them to the complex information system. Step 4.1: Retrieve supplementary information related to the final output of the current task from the knowledge base, vector library, and graph library of the data platform application through the knowledge retrieval unit; Step 4.2: The result fusion unit uses a large language model to integrate the output results of each node tool with the supplementary information obtained from the retrieval, generates structured decision results according to a predefined result template, and returns them to the complex information system through a standardized interface.

[0065] Step 5: The feedback optimization module in the intelligent hub will feed back the data of the entire task process and store it in the multimodal database to support the iterative optimization of the intelligent hub; The feedback optimization module feeds back the entire task process data and stores it in the multimodal database. The processing steps are as follows: Step 5.1: The process data and result data of the entire task are fed back through the feedback channel via the data feedback unit; Step 5.2: Perform ETL cleaning on the returned data through the data processing unit, convert the cleaned data into a vectorized representation and add classification labels to generate training samples in a standard format; Step 5.3: Store the training samples in the standard format into the vector library, knowledge base, and graph library of the multimodal database, respectively; Step 5.4: Based on the accumulated training samples in the vector library, knowledge base, and graph library, the model iteration unit periodically triggers the fine-tuning training of the large language model, and at the same time updates the embedding representation of the dual graph in the task decomposition module and the differentiable sortable feature weights in the tool recommendation and invocation module, so as to realize the optimization and iteration of the strategies of each module of the intelligent hub.

[0066] Before step 1, the process also includes building a toolset within the data platform application, with the following steps: The metadata of various tools within the data platform application is uniformly registered to the tool registration center. The metadata includes at least the tool's name, functional semantic description, tags, input parameter specifications, and output format specifications. Based on the input parameter specifications and output format specifications stored in the tool registry, the operation interface of each type of tool is encapsulated into a structured API interface through the interface encapsulation unit, providing a unified calling specification to shield the heterogeneity of each tool. During the process of the intelligent hub calling various tools to execute tasks, the continuous learning unit collects logs and performance data generated by the tools and feeds the logs and performance data back to the tool training and optimization process to achieve continuous improvement of tool capabilities.

[0067] This application also discloses a storage medium, which is a computer-readable storage medium storing a computer program thereon. When the computer program is executed, it can perform the steps of the aforementioned method. The storage medium may include read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0068] Application scenarios of this invention: Scenario 1: Application of Intelligent Scheduling System The intelligent hub is connected to the enterprise-level scheduling system to receive natural language instructions from the scheduler and automatically call tools such as prediction models, path planning models, and resource allocation models in the data platform to generate the optimal scheduling plan.

[0069] Scenario 2: Application of intelligent government systems The intelligent hub is connected to the government service system to receive natural language requests submitted by citizens, automatically call up tools such as policy matching models, process orchestration models, and document generation models in the data platform, and output processing suggestions.

[0070] Scenario 3: Application of Intelligent Medical Assistance Systems By integrating the intelligent central processing unit into the medical decision support system, it receives the patient's description of the condition input by the doctor and automatically calls upon the image analysis model, medical record retrieval model, and clinical guideline matching tool in the data platform to assist the doctor in making diagnostic decisions.

Claims

1. A method for integrating complex information systems and data platform applications based on large language model intelligent agents, characterized in that, include: Step 1: An intelligent hub is deployed between the complex information system and the data platform application. The intent understanding module in the intelligent hub performs semantic parsing and event element extraction on the complex tasks in natural language form issued by the complex information system, and converts the complex tasks in natural language form into structured task data. Step 2: The task decomposition module in the intelligent hub receives the structured task data in Step 1, performs macro-node arrangement based on the preset decision chain template, and generates sub-task sequences within each node using the dual-graph knowledge enhancement mechanism. Step 3: The tool recommendation and invocation module in the intelligent hub receives the sub-task sequence in step 2, recalls candidate tools from the tool set of the data platform application, sorts the candidate tools, and triggers the execution of the optimal tool by function call. Step 4: The integrated service module in the intelligent hub receives the execution results of the optimal tools corresponding to each node, integrates the execution results of the optimal tools corresponding to each node with external knowledge, generates structured decision results, and returns them to the complex information system. Step 5: The feedback optimization module in the intelligent hub will feed back the data of the entire task process and store it in the multimodal database of the data platform application to support the iterative optimization of the intelligent hub.

2. The method for integrating complex information systems and data platform applications based on large language model intelligent agents according to claim 1, characterized in that, In step 1, the intent understanding module processes the following steps: Step 1.1: Use the large language model in the semantic parsing unit to perform semantic parsing on the complex task in the form of natural language, and extract event elements and constraints. The event elements include at least time, location, task type, and event level. Step 1.2: Use the dialogue state tracking unit to generate a state vector based on the multi-turn dialogue history encoding. When event elements are missing, ask the user follow-up questions through guided dialogue. Step 1.3: Using the structured output unit, organize the extracted elements into structured task data according to the predefined structured output format.

3. The method for integrating complex information systems and data platforms based on large language model intelligent agents according to claim 1, characterized in that, In step 2, the task decomposition module processes the following steps: Step 2.1: Retrieve candidate decision chains from a pre-set decision chain template library based on semantic vector matching using the decision chain matching unit; Step 2.2: The node sequence defined by the selected decision chain is used as the macro framework for task execution by the macro node orchestration unit. The node sequence includes linear serial processing stages. Step 2.3: By calling the dual-graph knowledge enhancement mechanism in each node through the micro-node decomposition unit, the macro-task of the current node is decomposed into a sequence of specific sub-tasks that can be directly executed by the tool.

4. The method for integrating complex information systems and data platform applications based on large language model intelligent agents according to claim 3, characterized in that, In step 2.3, the processing steps of the dual-graph knowledge enhancement mechanism are as follows: Step 2.3.1: Call the large language model and combine it with semantic similarity calculation to extract the set of key entities from the task text of the current node; Step 2.3.2: Retrieve temporal dependency and logical constraint triples from the constraint relationship graph, retrieve semantic association triples from the domain knowledge graph, and aggregate them into structured knowledge fragments by D-hop adjacency subgraph; Step 2.3.3: Retrieve text fragments from the unstructured knowledge base of the data platform application using retrieval enhancement generation technology, and then splice and merge the structured knowledge fragments with the unstructured knowledge fragments to generate a knowledge-enhanced document; Step 2.3.4: Input the knowledge-enhanced document as prompt words into the large language model, guide the large language model to generate a thought chain corresponding to the logical structure of the decision chain nodes, and output a specific sub-task sequence.

5. The method for integrating complex information systems and data platform applications based on large language model intelligent agents according to claim 1, characterized in that, In step 3, the tool recommendation and invocation module processing steps are as follows: Step 3.1: Using a semantic encoder, the semantic recall unit encodes the subtask text and the metadata description of each tool in the toolset, sorts them in descending order of cosine similarity, and recalls the Top-K candidate tools from the toolset. Step 3.2: Construct a multidimensional feature vector for the candidate tools through a differentiable sorting unit and input it into the LapSum differentiable sorting operator. Output a comprehensive sorting value and select the optimal tool in descending order. Step 3.3: The large language model automatically fills in the parameters according to the input specifications of the selected optimal tool and the subtask context through the request generation unit, and generates a call request that conforms to the predefined JSON pattern; Step 3.4: Parse the call request through the interface triggering unit, trigger the execution of the corresponding microservice through the unified API gateway of the toolset, and return the standardized execution result; For each subtask, repeat sub-steps 3.1 to 3.4 until all subtasks in the current decision chain are assigned and the corresponding optimal tool is executed, thus obtaining the set of optimal tool execution results for each node.

6. The method for integrating complex information systems and data platforms based on large language model intelligent agents according to claim 1, characterized in that, In step 4, the processing steps of the integrated service module are as follows: Step 4.1: Retrieve supplementary information related to the final output of the current task from the knowledge base, vector library, and graph library of the data platform application through the knowledge retrieval unit; Step 4.2: The result fusion unit uses a large language model to integrate the output results of each node tool with the supplementary information obtained from the retrieval, generates structured decision results according to a predefined result template, and returns them to the complex information system through a standardized interface.

7. The method for integrating complex information systems and data platforms based on large language model intelligent agents according to claim 1, characterized in that, The feedback optimization module feeds back the entire task process data and stores it in the multimodal database. The processing steps are as follows: Step 5.1: The process data and result data of the entire task are fed back through the feedback channel via the data feedback unit; Step 5.2: Perform ETL cleaning on the returned data through the data processing unit, convert the cleaned data into a vectorized representation and add classification labels to generate training samples in a standard format; Step 5.3: Store the training samples in the standard format into the vector library, knowledge base, and graph library of the multimodal database, respectively; Step 5.4: Based on the accumulated training samples in the vector library, knowledge base, and graph library, the model iteration unit periodically triggers the fine-tuning training of the large language model, and at the same time updates the embedding representation of the dual graph in the task decomposition module and the differentiable sortable feature weights in the tool recommendation and invocation module, so as to realize the optimization and iteration of the strategies of each module of the intelligent hub.

8. The method for integrating complex information systems and data platforms based on large language model intelligent agents according to claim 1, characterized in that, Before step 1, the process also includes building a toolset within the data platform application, with the following steps: The metadata of various tools within the data platform application is uniformly registered to the tool registration center. The metadata includes at least the tool's name, functional semantic description, tags, input parameter specifications, and output format specifications. Based on the input parameter specifications and output format specifications stored in the tool registry, the operation interface of each type of tool is encapsulated into a structured API interface through the interface encapsulation unit, providing a unified calling specification to shield the heterogeneity of each tool. During the process of the intelligent hub calling various tools to execute tasks, the continuous learning unit collects logs and performance data generated by the tools and feeds the logs and performance data back to the tool training and optimization process to achieve continuous improvement of tool capabilities.

9. A complex information system and data platform application fusion system based on a large language model intelligent agent, characterized in that, It includes a complex information system, an intelligent hub, and a data platform application. The intelligent hub is deployed between the complex information system and the data platform application and includes an intent understanding module, a task decomposition module, a tool recommendation and invocation module, a comprehensive service module, and a feedback optimization module. The intent understanding module is used to perform semantic parsing and event element extraction on complex tasks in natural language form issued by the complex information system, and to convert the complex tasks in natural language form into structured task data. The task decomposition module is used to receive structured task data obtained by the intent understanding module, perform macro-node arrangement based on a preset decision chain template, and generate sub-task sequences within each node using a dual-graph knowledge enhancement mechanism. The tool recommendation and invocation module is used to receive the sub-task sequence obtained by the task decomposition module, recall candidate tools from the tool set of the data platform application, sort the candidate tools, and trigger the execution of the optimal tool in the form of a function call. The integrated service module is used to receive the optimal tool execution results corresponding to each node obtained by the tool recommendation and invocation module, integrate and organize the optimal tool execution results corresponding to each node with external knowledge, generate structured decision results, and return them to the complex information system. The feedback optimization module is used to feed back the data of the entire task process and store it in the multimodal database of the data platform application to support the iterative optimization of the intelligent hub.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by the processor to perform the steps of the fusion method according to any one of claims 1 to 8.

Images