An enterprise intelligent middleware system based on natural language and visual orchestration

By utilizing the process generation, model scheduling, and tenant management modules of the enterprise intelligent middleware system, the problems of inconsistent interfaces and high operation and maintenance costs in the construction of enterprise AI business have been solved, and end-to-end integrated management and compliance auditing have been achieved.

CN122837813BActive Publication Date: 2026-11-10TAIYI (SANYA) TECHNOLOGY CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611339168.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-10
Estimated Expiration
2046-09-01

Smart Images

  • Figure CN122837813B_ABST
    Figure CN122837813B_ABST
Patent Text Reader

Abstract

This invention discloses an enterprise intelligent middleware system based on natural language and visual orchestration, relating to the field of enterprise-level artificial intelligence application management. The system includes: a process generation module that parses business intent into standardized workflow topologies; a model scheduling module with a built-in model selection algorithm that uniformly encapsulates multi-vendor model interfaces; an intelligent routing module that dynamically matches the best intelligent agent to execute dialogues and complex business processes based on an intelligent matching algorithm; a capability opening module that exposes AI resources externally through standard long connections; and a tenant management module that constructs a four-level data isolation and fine-grained permission system, providing a natural language operation and maintenance entry point. This invention establishes a complete closed loop encompassing natural language understanding, workflow orchestration, model scheduling, intelligent routing, and open invocation, reducing the complexity of enterprise AI application development and operation and maintenance, improving resource utilization efficiency and system observability in multi-tenant scenarios, and is suitable for the rapid construction and large-scale deployment of enterprise-level AI middleware.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of enterprise-level artificial intelligence application management, specifically to an enterprise intelligent middleware system based on natural language and visual orchestration. Background Technology

[0002] As digital transformation continues, various enterprises are beginning to implement AI on a large scale, with large models, intelligent agents, and automated business processes becoming the core carriers of enterprise digital construction. The market offers numerous large model products from different vendors, with inconsistent interface specifications, calling logic, and input / output formats. Enterprises need to individually interface with each type of model service when building AI businesses, resulting in cumbersome integration processes and high maintenance costs. Simultaneously, enterprise business needs are complex and ever-changing, with many business processes requiring multi-step reasoning, knowledge base retrieval, and multi-round model calls. Traditional code development models require users to have professional programming skills, making it impossible for business personnel to independently build business logic. External business systems also need to connect to internal AI resources, but there is a lack of standardized, secure, and controllable external access channels. The scenario of multiple enterprises sharing platforms is gradually increasing, with a lack of comprehensive isolation and control measures for data, resources, and permissions across different enterprises. Various operational data and operation records are stored in a scattered manner, making unified auditing and resource usage coordination difficult. The industry lacks an integrated middleware solution that can integrate process orchestration, model scheduling, dialogue agents, external capability output, and full-link tenant management.

[0003] Existing enterprise AI deployment technologies suffer from several unavoidable practical shortcomings. At the business process level, they rely solely on coding business flow logic, lacking natural language translation capabilities, visual editing, and full-process verification mechanisms. This makes process logic prone to loops and isolated issues, and the absence of standardized version releases and runtime backtracking capabilities significantly complicates troubleshooting. At the model integration level, adding new model vendors requires modifying underlying service code, hindering lightweight hot integration. The lack of automatic matching mechanisms for business scenarios necessitates manual model specification, making it difficult to balance service effectiveness and usage costs. At the dialogue interaction level, single matching logic cannot adapt to different business scenarios. High-concurrency scenarios result in high computational overhead, and high-precision business logic lacks deep semantic matching capabilities. Dialogues cannot be linked to automated business processes, only achieving basic question-and-answer functionality. At the tenant management level, the lack of a layered data isolation architecture leads to risks of unauthorized data access when multiple enterprises share the platform. Permission and resource quota control is coarse-grained, and maintenance relies on specialized technical personnel. There are no lightweight natural language maintenance channels, and the full-link operation logs are incomplete, failing to meet industry compliance audit requirements. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an enterprise intelligent middleware system based on natural language and visual orchestration. The process generation module parses business intent into standardized workflow topologies, supporting visual orchestration and version management; the model scheduling module has a built-in model selection algorithm, uniformly encapsulates multi-vendor model interfaces, and achieves optimal adaptation between scenarios and models; the intelligent routing module relies on the intelligent matching algorithm to dynamically match the best intelligent agent to execute dialogues and complex business processes; the capability opening module exposes AI resources externally through standard long connections, supporting external client discovery and invocation; the tenant management module constructs a four-level data isolation and fine-grained permission system, provides a natural language operation and maintenance entry point, and achieves unified management and audit traceability across the entire chain.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an enterprise intelligent middleware system based on natural language and visual orchestration, the system comprising:

[0006] Workflow generation module: Receives natural language business requirements, parses the natural language requirements into node specification mappings, generates a standardized directed acyclic graph structure workflow topology after topology verification, provides a visual drag-and-drop editing interface, supports workflow version management and runtime tracking, outputs standardized workflow data, and synchronizes all resource attribute data to tenant management.

[0007] Model scheduling module: Receives standardized workflow data, has a built-in model selection algorithm, uniformly stores all model metadata, automatically selects and adapts models based on the scene identifiers passed from upstream, uniformly encapsulates the model interfaces of different vendors to provide a unified computing power call entry point, records all call usage data and synchronizes it to tenant management.

[0008] Intelligent routing module: Relying on a unified computing power entry point, with a built-in intelligent matching algorithm, it receives user dialogue requests, performs matching and filtering on all available intelligent agents within the tenant, selects the optimal intelligent agent to execute dialogue tasks, calls the workflow generated by the process generation module to complete complex business, and synchronizes the dialogue operation log to the tenant management system.

[0009] Capability Opening Module: Based on the standard long connection protocol and MCP standard protocol, it opens up all AI business resources of the tenant to the outside world. It has four types of standard connectors built-in: HTTP / REST, MCP Server, RPC, and database query. It supports resource discovery and invocation by external clients, and links the process generation module, model scheduling module, and intelligent routing module to execute corresponding capability operations, and generates full audit logs that are synchronized to the tenant management system.

[0010] Tenant Management Module: Covers all aspects of the process generation module, model scheduling module, intelligent routing module, and capability opening module. It supports a four-level organizational structure, manages tenants' ability to subscribe to AI resources on the platform and their fine-grained permissions, provides a natural language operation and maintenance entry point, issues permissions and quota rules to the process generation module, model scheduling module, intelligent routing module, and capability opening module, and aggregates the platform's operational data to achieve unified management and control across the entire chain.

[0011] Furthermore, the natural language business requirements received by the process generation module include at least one of six categories of information: business scenario objectives, data flow logic, knowledge base call requirements, model output format constraints, branch judgment conditions, and result output specifications. The process generation module retrieves the built-in complete node specifications, workflow topology hard constraints, and structured prompt word templates, concatenates the requirement information with the templates, and sends them to the model scheduling module for inference execution. After the model returns the original topology JSON data, the process generation module sequentially performs node type whitelist verification, topology island and loop detection, and start and end node integrity verification. When an anomaly is found during verification, it is automatically corrected. After the verification passes, the canvas node coordinates are automatically arranged and the default values ​​of the inference parameters are completed. Standardized workflow data is generated and pushed to the visualization interface for rendering. It supports personnel to add or delete nodes, adjust connections, and modify the configuration of each node. After editing, the draft is stored, and the final version is generated after approval.

[0012] Furthermore, the workflow generation module incorporates a version management and release approval mechanism. Workflows edited by users with editing permissions are stored as drafts by default, and can be modified multiple times to overwrite the original draft content. After confirming the workflow logic, a release application is submitted. Once approved by the approver, an unmodifiable official release version is generated, with each official release version corresponding to an independent and stable calling address. Each time the workflow is invoked, a complete running record and a single-node execution record are generated, storing the input content, output content, execution status, and running time of each node. The running data is synchronously pushed to the tenant management system for usage statistics and anomaly investigation. Users can trace the complete execution chain through a visual interface, locate abnormal nodes in the workflow, and carry out optimization and adjustments.

[0013] Furthermore, the model scheduling module's built-in model selection algorithm is used for scene and model fit scoring. After receiving the scene identifier, it first filters a set of candidate models that meet the basic capability requirements, and then performs a scoring operation on all models in the set. The operation expression is: In the formula This represents the fit score of the m-th single candidate model for the current scene. The higher the value, the stronger the fit. The weight of the capability matching dimension ranges from 0.4 to 0.6. This represents the capability matching factor, which takes a value of 1 when the model fully covers the capabilities required for the scenario, and a value of 0 when there are missing capabilities. This represents the weight of the performance dimension, with a value ranging from 0.2 to 0.3; The performance factor is calculated from the historical average response time of the model. The shorter the response time, the higher the value. The value ranges from 0 to 1. This represents the weight of the cost control dimension, with a value ranging from 0.2 to 0.3. This represents the cost factor, calculated from the cost of the model unit. The lower the cost, the higher the value. The value ranges from 0 to 1. + + =1; the model with the highest score is selected to perform the current task.

[0014] Furthermore, the model scheduling module stores all model metadata. Each piece of metadata includes the model's vendor, model category, functional support status, vector dimension, and input / output modal attributes. It supports dynamically adding model vendor configurations; simply filling in the vendor's interface address, authentication key, and timeout threshold is sufficient for access. Configuration changes take effect in real time without requiring a service restart. For established long-connection sessions, configuration changes do not affect existing sessions but take effect when creating new sessions. It has built-in unified request conversion logic. After receiving standardized call parameters from upstream, it automatically converts them into the request format required by the corresponding vendor. After obtaining the vendor's return results, it uniformly organizes them into a standard format and sends them back to upstream, completely shielding the interface differences between different vendors. Each call records the calling subject, model name, resource consumption, and execution time data, which is synchronized to tenant management for quota control and usage statistics.

[0015] Furthermore, the intelligent routing module's built-in intelligent matching algorithm is used to score the matching degree between user queries and intelligent agents. After receiving the user's query text, it retrieves the description text and tag set of all enabled intelligent agents within the tenant, and performs matching degree calculation on each intelligent agent one by one. The calculation expression is as follows: In the formula For the first The higher the score, the more suitable the agent is for handling the current query. The semantic similarity factor is used to compare the user's query text. With agent description text After being encoded into vectors using the Embedding model, the cosine similarity between the two vectors is calculated, with the value ranging from [value range missing]. to A larger value indicates that the semantics are closer. As a capability matching factor, the Jaccard coefficient is used to calculate the set of capabilities required by the user query and the agent. Capabilities The degree of matching, i.e., the ratio of the intersection size to the union size, has a range of values. to ; As a performance factor, based on intelligent agents Historical average response time parameter The conversion shows that the shorter the time, the higher the value of this factor, and the range of values... to The conversion method is as follows Or a normalized mapping can be used, where Indicates the first The historical average response time (in seconds) for each agent. The attenuation coefficient; As a cost factor, determined by the agent The model unit call cost parameter on which it depends In other words, the lower the cost, the higher the value of this factor, and the range of values... to The conversion method is as follows ,in For the first The cost of a single invocation of the model relied upon by an agent. and These are the highest and lowest costs among all available models within the current tenant; For strict matching factors, it has built-in rule sets for eight industry sectors: finance, law, education, healthcare, manufacturing, retail, government affairs, and office. (If querying...) Identified as belonging to a specific professional field, and the intelligent agent Description If the same field is explicitly marked, then the value is [value]. Otherwise ; The weight coefficients for the five factors mentioned above are determined by the business team through pairwise comparisons of the importance of each factor using the analytic hierarchy process, and satisfy the following conditions: The range of values ​​for each weight is dynamically adjusted according to the business scenario. When the comprehensive score of multiple candidate agents is... When the maximum difference between the scores is less than the preset threshold of 0.05, the system automatically triggers the large language model to rearrange the intent of agents with similar scores. This means that the descriptions, labels, and original user queries of these agents are resubmitted to the large model for secondary semantic adjudication. Finally, the large model outputs the unique optimal matching agent as the final routing result.

[0016] Furthermore, the intelligent routing module includes two sets of matching execution strategies, which tenant administrators can independently select the effective strategy according to the business scenario. The first set is a semantic priority strategy, which calls the intelligent matching algorithm to perform semantic matching operations on all intelligent agents, adapting to business scenarios where the accuracy of the answer reaches a preset threshold, where the preset threshold is a comprehensive matching score ≥ 0.75. The second set is a fast matching strategy, which, based on the pre-stored keyword-agent mapping relationship, compares the keywords contained in the query text to directly hit the corresponding intelligent agent, skipping the vectorization operation step, and adapting to high-concurrency, low-latency business scenarios.

[0017] Furthermore, after the intelligent routing module selects the optimal agent, it executes the dialogue task. First, it reads the exclusive prompt words bound to the agent, the associated knowledge base retrieval rules, and the basic inference parameters. Relying on the unified computing power entry point, it initiates a vector retrieval, and splices the retrieved knowledge base fragments and the user's original dialogue request into a complete inference message, which is then sent to the large model to complete the basic question-and-answer output. When the optimal agent is configured to bind a business process identifier, the composite business execution logic is triggered. According to the business process identifier bound to the agent, the corresponding published workflow definition is retrieved, and the current dialogue context variables and tenant identity information are transmitted to the workflow generation module. This drives the workflow generation module to start the corresponding standardized workflow instance. The computing power requirements of each node in the workflow continuously call the model scheduling module to support the computation. After the entire workflow is fully executed, the process output results are summarized and integrated with the basic question-and-answer content to generate the final dialogue response. The matching records, model consumption, and process call records generated throughout the dialogue process are synchronously pushed to the tenant management module for archiving.

[0018] Furthermore, the exclusive prompt words include agent identity positioning, business response specifications, answer format constraints, and prohibited output content definition; the associated knowledge base retrieval rules include retrieval matching threshold, number of documents returned in a single instance, document filtering tags, and retrieval field range configuration; the basic inference parameters include model temperature value, maximum output character length in a single instance, tool call switch, and structured output mandatory identifier.

[0019] Furthermore, the capability opening module exposes tenant-internal AI resources based on a standard long-connection protocol, while also supporting MCP (ModelContextProtocol) standard protocol access. It provides SSE long-connection endpoints for standard MCP clients such as Cursor and Windsurf to access, and uses the JSON-RPC protocol for tool discovery and invocation. It supports access to external MCP servers via both HTTP and STDIO transport protocols, automatically discovering and synchronizing the input schema and parameter definitions of external tools. It also includes a built-in connector registry for registering HTTP / REST connectors and MCPs. There are four types of standard connectors: Server connector, RPC connector, and database query connector. Each connector is pre-configured with authentication methods, permission boundaries, data anonymization strategies, and risk control rules. When an external client initiates a connection request, it must carry a unique identity credential. After successful credential verification, a persistent long-lived connection session is established, and the connection is actively disconnected when the credential expires. When a client initiates a resource list query request, it automatically scans all enabled agents, knowledge bases, skills, and published workflows within the current tenant, encapsulates each resource into a standard tool description, and returns it via a long-lived connection. When a client initiates a resource call request, it parses the type of the called resource and distributes it to the process generation module, model scheduling module, or intelligent routing module to execute the corresponding operation. The execution result is returned in segments via a long-lived connection. Each call generates an audit log, which includes the caller identifier, resource name, input / output summary, and execution time data, and is synchronized to the tenant management and retention database.

[0020] Furthermore, the tenant management module constructs a four-level organizational structure, with the levels being platform, tenant enterprise, internal department, and end user. All business resources are bound to a unique tenant identifier, and all requests carry tenant context, ensuring complete data isolation between different tenants and preventing cross-authority access. The tenant management module has multiple subscription versions built-in, each bound to a corresponding set of available capabilities. Before executing a resource creation operation, the current tenant's subscription range is verified. This subscription range includes five quantity-based quota constraints: the maximum number of workflow creations, the maximum number of intelligent agents activated, the maximum knowledge base storage capacity, the number of concurrent connections to MCP external tools, and the multimodal model call quota. Operations exceeding these limits are directly intercepted. A fine-grained role-based permission system is configured, supporting custom roles and granting menu access and resource operation permissions, including various permissions such as operation and maintenance dialogue permissions. Different levels of management personnel have different operation ranges. The tenant management module distributes all permissions and quota rules to the process generation module, model scheduling module, intelligent routing module, and capability opening module in real time as the verification basis before operation execution.

[0021] Furthermore, the tenant management module provides a dedicated natural language operation and maintenance (O&M) entry point, accessible only to administrators with O&M privileges. After administrators input natural language O&M commands, the linkage model scheduling module parses the commands, identifies the operation type and corresponding parameters, and covers resource creation, quota adjustment, permission change, usage query, and service start / stop. After parsing, the corresponding internal interface is called to execute the actual operation. The operation risk is automatically determined based on the risk level definition; high-risk operations trigger a secondary confirmation process, and execution only proceeds after confirmation. High-risk operations include at least resource deletion, quota reduction, and permission revocation. After the operation is completed, the execution result is compiled into natural language text and fed back to the administrators. All O&M operations are logged, recording the operator, operation time, operation content, and status before and after changes, meeting audit and traceability requirements.

[0022] Furthermore, the risk level is defined into three levels based on the degree to which the operation causes irreversible impact on tenant resources, platform permissions, and call quotas. Level 1 is low risk, including read-only operations such as usage query and configuration viewing; Level 2 is medium risk, including new intelligent agents, knowledge base expansion, and call quota increase operations; and Level 3 is high risk, including three types of irreversible change operations such as resource deletion, quota reduction, and permission revocation.

[0023] Compared with existing technologies, this enterprise intelligent middleware system based on natural language and visual orchestration has the following advantages:

[0024] I. This invention integrates natural language parsing, visual workflow orchestration, and unified scheduling capabilities for multi-vendor models through a layered modular architecture. It relies on the collaborative operation of four major business modules: independent workflow generation, model scheduling, intelligent routing, and capability openness, coupled with a unified tenant management system to achieve end-to-end coordination. It can automatically convert business personnel's natural language requirements into standardized workflow topologies, with a visual drag-and-drop interface lowering the barrier to workflow development. A built-in multi-layer topology verification mechanism automatically corrects logical defects, fully covering the entire lifecycle from draft editing, approval and release, to runtime backtracking. It uniformly encapsulates the differences in interfaces of various large models, automatically matching the optimal inference model based on scenario adaptation algorithms, eliminating the need for manual model resource switching. Multi-dimensional usage data is synchronized to the management center in real time, enabling on-demand allocation of computing resources and reducing the cost of building and maintaining enterprise AI applications.

[0025] II. This invention improves the dialogue and resource openness system through a dual-layer intelligent matching strategy and hierarchical tenant isolation and control. It provides two sets of intelligent agent matching logic to address different concurrency and accuracy requirements, balancing high-concurrency response speed with the accuracy of complex business question-and-answer. It supports the execution of complex business processes through a visualized workflow linked to dialogue scenarios, connecting knowledge base retrieval, model reasoning, and process calculation to form a complete business loop. Externally, it opens up the entire platform's AI capabilities based on standard long connections, and fully retains end-to-end audit logs to meet compliance and traceability requirements. The platform establishes a four-level data isolation architecture and a hierarchical permission quota system, coupled with a natural language operation and maintenance channel. It performs hierarchical verification based on the risk level of operation impact, ensuring that data from multiple enterprise tenants does not interfere with each other while simplifying the platform's operation and maintenance threshold, forming a refined, secure, and controllable enterprise AI resource management system.

[0026] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0028] Figure 1 A flowchart of an enterprise intelligent middleware system based on natural language and visual orchestration;

[0029] Figure 2 This is a data transmission diagram of an enterprise intelligent middleware system based on natural language and visual orchestration.

[0030] Figure 3 This is a flowchart of the model scheduling module of the present invention. Detailed Implementation

[0031] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0032] Example 1:

[0033] This embodiment uses a heavy equipment manufacturing enterprise as an independent tenant, and relies on this system to build an automated procurement application and approval system covering all subsidiaries, implementing AI-driven procurement operations throughout the entire process. Figure 1As shown, the complete implementation process is as follows:

[0034] After logging into the platform, enterprise tenant administrators access the tenant management module to complete the full-dimensional basic system configuration. Information is entered in a hierarchical manner, strictly following the four-level organizational structure of platform, tenant enterprise, internal department, and end-user. First, the unique identifier of the enterprise's main tenant is registered. Then, departmental levels are created sequentially: group purchasing headquarters branch, purchasing departments of each production plant, finance department, and audit department. Finally, employee accounts for five categories of positions—purchasing handlers, department heads, financial reviewers, auditors, and IT maintenance personnel—are imported in batches. Custom roles are defined for different positions. Purchasing handlers are granted workflow editing and draft modification permissions; finance managers are granted process approval and release permissions; auditors are only granted permission to view process execution records; and IT maintenance personnel are granted dedicated natural language access to the maintenance interface. The administrator selects a subscription version that matches the group's business volume and checks each of the five quota indicators: maximum number of workflow creations, maximum number of intelligent agents enabled, maximum knowledge base storage capacity, concurrent connections to MCP external tools, and multimodal model call quota. The administrator confirms that the values ​​can handle the group's annual processing needs of over 100,000 purchase orders. After all configurations are completed, the tenant management module automatically distributes the permissions and quota rules to the process generation module, model scheduling module, intelligent routing module, and capability opening module in real time. Before all business operations are executed, the distributed rules are automatically matched and verified. Operations that exceed the quota or lack corresponding permissions are directly blocked, thus avoiding the problems of unauthorized resource creation and unauthorized operations from the source.

[0035] Group procurement personnel with editing permissions enter the visual drag-and-drop editing interface of the process generation module. They input complete business requirements into the natural language input box, covering six categories of information: procurement application business scenario objectives, procurement document data flow logic, procurement compliance system knowledge base call requirements, model output structured form constraints, procurement amount hierarchical approval branch judgment conditions, and approval / rejection result output specifications. The process generation module retrieves locally built-in complete node specifications, workflow topology hard constraints, and standardized structured prompt word templates, concatenating the input requirements with the templates to form an inference message, which is then transmitted to the model scheduling module for topology structure inference. The model scheduling module receives the procurement approval-specific scenario identifier, activates the built-in model selection algorithm to filter the appropriate model set, and after completing the topology inference, sends the original topology JSON data back to the process generation module. The model selection algorithm's calculation expression is... In the formula This represents the fit score of the m-th single candidate model for the current scene. The higher the value, the stronger the fit. The weight of the capability matching dimension ranges from 0.4 to 0.6. This represents the capability matching factor, which takes a value of 1 when the model fully covers the capabilities required for the scenario, and a value of 0 when there are missing capabilities. This represents the weight of the performance dimension, with a value ranging from 0.2 to 0.3; The performance factor is calculated from the historical average response time of the model. The shorter the response time, the higher the value. The value ranges from 0 to 1. This represents the weight of the cost control dimension, with a value ranging from 0.2 to 0.3. This represents the cost factor, calculated from the cost of the model unit. The lower the cost, the higher the value. The value ranges from 0 to 1. + + =1; The model with the highest score is selected to execute the current task. The process generation module performs node type whitelist verification, topology island and loop detection, and start and end node integrity verification in a fixed order. Detected logical anomalies are automatically corrected by the system. After all verifications pass, the system automatically arranges the coordinates of each node in the canvas, synchronously completes the default values ​​of the inference parameters for each node, and generates standardized workflow data which is pushed to the visualization interface for real-time rendering. Personnel can drag and drop to adjust node arrangement, add or remove compliance verification, hierarchical approval, and other nodes in the interface, and modify the knowledge base retrieval parameters for each node. After adjustment, the draft is saved, and the draft supports dozens of iterations and can overwrite the original content. After confirming that the procurement approval logic has no loopholes, a release application is submitted. The finance manager verifies the compliance terms of the process online. After approval, an unmodifiable official release version is generated. Each official version is bound to an independent and stable call address, and external systems can specify the version to ensure business stability when calling it. The final procurement approval workflow includes six core nodes: application information submission, automatic verification of procurement compliance, hierarchical branching of amounts, approval by branch plant supervisors, review by group finance, and notification of approval results. It fully covers the entire document processing chain for small-amount parts procurement and large-scale equipment bidding procurement.

[0036] During the formal production and operation phase of the procurement approval workflow, the model scheduling module uniformly manages the metadata of all vendor-accessed models. Each metadata entry fully retains information including the vendor, model category, functional support status, vector dimension, and input / output modal attributes. Subsequent additions of invoice recognition and text summarization models from other vendors do not require restarting the middleware service; only the vendor's interface address, authentication key, and timeout threshold need to be filled in to complete the integration. Configuration changes only affect newly established long-lived connections; existing connections remain unaffected, ensuring uninterrupted business operations. Different nodes, such as document information extraction, compliance clause comparison, and approval opinion generation, will send unique scenario identifiers to the model scheduling module. The module then initiates a model selection algorithm to score the suitability of candidate models, selecting the optimal model across all dimensions to undertake the corresponding node's computational tasks. The module incorporates unified request conversion logic, uniformly encapsulating multi-vendor interfaces to shield interface differences and provide a unified computing power entry point. After receiving standardized upstream call parameters, it automatically adapts to the differentiated interface formats of various vendors. After vendors return computation results, the data is uniformly formatted and sent back to the upstream, completely shielding the integration costs caused by inconsistent interface formats among multiple vendors. Each model call will retain the calling entity, model name, resource consumption, and execution time details, and push them synchronously to the tenant management module for persistent storage. This is used to calculate the total model usage on a monthly basis and monitor the quota consumption progress. When the usage is close to the quota limit, the system will automatically push an early warning.

[0037] When employees within the group submit inquiries about procurement policies or initiate online procurement requests through the company's internal AI dialogue portal, the intelligent routing module receives user dialogue requests in real time and retrieves all enabled intelligent agents within the tenant to perform matching and screening. Tenant administrators can independently switch matching execution strategies monthly based on business volume. For business scenarios such as interpretations of large-scale equipment procurement policies and inquiries about bidding and tendering regulations, where the accuracy of answers reaches a preset threshold, a semantic priority strategy is activated. This strategy invokes the intelligent matching algorithm to complete semantic matching calculations for all intelligent agents. The intelligent matching algorithm's calculation expression is: In the formula For the first The combined score of the matching degree between each candidate agent and the current user query; It is a semantic similarity factor; Ability matching factor; For performance factors; Cost factor; For strict domain matching factors; This represents the user's query text. Indicates the first Description text of each agent Indicates the first The set of capabilities possessed by an intelligent agent Indicates the first Historical average response time parameter for each agent Indicates the first The unit call cost parameter of the model that each agent depends on; These are the weighting coefficients for the five factors mentioned above. For high-frequency and simple inquiries such as office supply requisition and small-amount reimbursement processes, a fast matching strategy is switched. Based on the pre-stored keyword-agent mapping relationship, the corresponding agent is directly matched, eliminating the vectorization calculation step and shortening the response time. After the intelligent routing module selects the procurement business agent with the highest matching degree, it reads the agent's bound exclusive prompt words, associated knowledge base retrieval rules, and basic inference parameters. Using the unified computing power entry provided by the model scheduling module, vector retrieval is performed. The retrieved procurement compliance knowledge base fragments and the user's original dialogue content are concatenated into a complete inference message and sent to the large model, outputting basic question-and-answer content. If a user uploads a purchase order to initiate a formal application, the procurement approval business process identifier bound to the intelligent agent will trigger the complex business execution logic. The system retrieves the complete definition of the published procurement approval workflow, synchronizes the dialogue context variables and tenant enterprise identity identifier to the process generation module, and drives the middle platform to generate a dedicated standardized workflow instance. During the instance's operation, the text extraction and compliance verification computing power requirements generated at each node continuously call the model scheduling module. After all nodes of the entire workflow have been executed, the system integrates the process approval results and basic Q&A content, generates a complete dialogue response including approval progress and pending items, and pushes it to the employee portal. The intelligent agent matching records, model resource consumption data, and workflow call records generated by the dialogue interaction are all synchronously archived to the tenant management module for easy subsequent business review and traceability.

[0038] The capability open module exposes tenant-internal AI resources based on a standard long-connection protocol, while also supporting MCP standard protocol access. It provides SSE long-connection endpoints for standard MCP clients such as Cursor and Windsurf to access, and uses the JSON-RPC protocol for tool discovery and invocation. It supports access to external MCP servers via HTTP and STDIO transport protocols, automatically discovering and synchronizing the input schema and parameter definitions of external tools. The system has a built-in connector registry, registering four types of standard connectors: HTTP / REST connectors, MCP Server connectors, RPC connectors, and database query connectors. Each connector is pre-configured with authentication methods, permission boundaries, data anonymization strategies, and risk control rules. The group's OA procurement system acts as an external client, establishing a persistent long-connection session upon access with a platform-issued exclusive identity credential. The module proactively disconnects after the credential expires, mitigating the risk of unauthorized continuous access. The OA system can initiate resource list query requests on demand. The module automatically scans all enabled agents, knowledge bases, tool skills, and published procurement workflows within the tenant, encapsulating each type of resource into a standard tool description and streaming it back via a long connection. When the OA system synchronizes offline paper procurement documents and pushes online approval pending tasks, the client initiates a resource call request. After identifying the resource type, the module distributes it to the process generation module, model scheduling module, or intelligent routing module for execution. Document extraction and approval workflow results are then streamed back to the OA front-end page in segments. Every external client call generates a complete audit log, recording the caller identifier, resource name, input / output summary, and overall execution time. This log is simultaneously pushed to the tenant management module for long-term retention, meeting the annual audit requirements of the group's audit department.

[0039] The tenant management module permeates all business processes within the middle platform, implementing a four-tier organizational structure encompassing the platform, tenant companies, internal departments, and end users. All group-wide procurement documents, processes, and intelligent agent resources are bound to a unique tenant identifier. Each business request carries tenant context information, preventing other manufacturing tenants from accessing the group's procurement data and eliminating the risk of cross-tenant data leakage. The tenant management module includes multiple subscription versions, each bound to a corresponding set of available capabilities. Every time the procurement workflow is invoked by the OA system or dialogue portal, the system automatically generates a complete execution record and single-node execution details, storing the input, output, execution status, and execution time of each node. All execution data is synchronously aggregated in the tenant management module's backend. Administrators can drag and drop to view the entire execution sequence through a visual interface, quickly locating abnormal nodes such as those experiencing verification failures or timeouts, and adjusting node knowledge base retrieval rules and model call parameters accordingly. The group's IT department operations personnel can only initiate operations through a dedicated natural language operations portal. After inputting natural language commands, the system, in conjunction with the model scheduling module, parses the command operation type and associated parameters. The scope of operations covers resource creation, quota adjustment, permission change, usage query, and service start / stop. The system automatically classifies operations into risk levels based on built-in risk level definitions. Level 1 low-risk operations, such as usage query and configuration viewing, are executed directly. Level 2 medium-risk operations, such as adding intelligent agents, expanding knowledge base storage capacity, and increasing model call quotas, take effect immediately. Level 3 high-risk operations, such as deleting historical procurement workflows, reducing multimodal call quotas, and revoking departmental approval permissions, require a secondary confirmation pop-up. The system will only execute the change after the operations personnel confirm that everything is correct. All operations are fully logged, recording the operator's account, operation time, operation content, and configuration status before and after the change. The audit department can export the logs at any time to complete compliance checks.

[0040] This equipment manufacturing enterprise procurement approval implementation example relies on five major modules to collaboratively achieve full-process digital procurement. The tenant management module, through a four-level organizational structure and hierarchical permission quota control, safeguards data security and operational standards. The process generation module automatically generates standardized procurement workflows using natural language to topology conversion and version approval mechanisms. The model scheduling module uses a model selection algorithm to achieve unified adaptation of multi-vendor models, shielding interface differences and providing stable computing power support. The intelligent routing module, with two matching strategies and intelligent matching algorithms, distinguishes between consultation and complex approval scenarios, linking workflows to complete the entire online procurement process. The capability opening module achieves data interoperability by connecting with the enterprise's existing OA system through long connections. After going live, document processing efficiency is improved, algorithm matching accuracy is stable, and the operation and maintenance risk interception mechanism effectively avoids human error, fully meeting the core business requirements of large manufacturing enterprises for procurement compliance, audit traceability, and multi-tenant isolation.

[0041] Example 2:

[0042] This embodiment uses a daily-use chain retail enterprise as an independent tenant, and builds an integrated AI business system based on this system, encompassing automatic handling of store anomalies and online consultation for store staff. Figure 2 As shown, the complete implementation process is as follows:

[0043] Retail enterprise headquarters administrators log in to the tenant management module to perform basic tenant configuration across the entire platform. They strictly adhere to a four-tier organizational structure—platform, tenant enterprise, internal departments, and end-users—to build a layered management system. First, they complete the registration of retail enterprise tenant entities and bind unique tenant identifiers. Then, they divide the structure into three levels: headquarters operations center, North China / East China / South China regional branches, and offline stores. They batch-enter four types of terminal accounts: headquarters operations specialists, store managers, frontline store staff, and headquarters IT maintenance personnel. Administrators define multiple operation roles. Headquarters operations personnel are granted workflow editing and intelligent agent creation and configuration permissions. Store staff nationwide are only granted access for dialogue consultation and anomaly reporting. IT maintenance personnel are separately assigned exclusive access to the natural language IT maintenance portal. The administrator selected the corresponding subscription tier based on the average daily business volume of 320 stores, and verified the five quotas item by item: the maximum number of workflow creations, the maximum number of intelligent agents activated, the maximum knowledge base storage capacity, the number of concurrent connections to MCP external tools, and the multimodal model call quota. It was confirmed that the quotas could handle the concurrent call demand during peak customer traffic periods. After all permissions and quotas were configured, the system automatically distributed the rules to the process generation module, model scheduling module, intelligent routing module, and capability opening module. Before all store reports, process editing, and model call operations were executed, the quotas and permissions were automatically verified. Requests exceeding the quotas or without operation permissions were blocked in real time to prevent unrestrained resource creation and unauthorized modification of processing procedures.

[0044] Headquarters operations specialists enter the visual drag-and-drop editing interface of the process generation module, and input complete store anomaly handling business requirements in the natural language input area. This includes six categories of information: the store operation anomaly handling scenario objectives, the store customer flow / equipment failure data flow logic, store operation standard knowledge base call requirements, anomaly handling solution output format constraints, fault level branch judgment conditions, and handling completion result feedback standards. The process generation module retrieves the system's built-in complete node specifications, workflow topology hard constraints, and standardized structured prompt word templates, concatenating the store business requirements with the templates to integrate them into an inference message, which is then transmitted to the model scheduling module for topology structure inference calculations. The model scheduling module receives the store anomaly handling-specific scenario identifier, activates the built-in model selection algorithm to filter the appropriate model set, and after completing the topology inference, sends the original topology JSON data back to the process generation module. The workflow generation module sequentially performs node type whitelist verification, topology island and loop detection, and start / end node integrity verification according to a fixed process. Detected logical breakpoints and loop issues are automatically corrected by the system. After all verification items pass, the system automatically arranges the coordinates of each node on the canvas, synchronously completes the default values ​​of each node's inference parameters, and generates standardized workflow data which is pushed to the visualization interface for real-time rendering. Operations specialists can drag and drop to adjust the node order, modify fault level judgment thresholds, and change the notification contact person on the interface. After adjustment, the draft is saved and can be iterated and modified multiple times to overwrite the original content. After confirming that the entire fault handling process logic is correct, a release application is submitted. The headquarters operations manager verifies the process's rationality online. Once approved, an unmodifiable official release version is generated. Each official handling workflow is bound to an independent and stable calling address. When stores use mobile terminals, they can specify the version to ensure consistent fault handling logic. The finalized workflow for handling store operational anomalies includes six core nodes: reporting of anomalies, automatic determination of fault level, matching of standardized handling solutions, notification of relevant regional managers, online verification of handling results, and closure and archiving of fault work orders. It comprehensively handles various operational anomaly work orders, including those related to bank equipment failures, product shortages, customer complaints, and store losses.

[0045] During the daily operation of the store anomaly handling workflow, the model scheduling module uniformly stores and manages the metadata of all vendor-accessed models. Each piece of metadata completely retains the model's vendor, model category, functional support status, vector dimension, and input / output modal attributes. Subsequent additions of vendor models for product recognition and customer complaint text classification do not require restarting the middleware service; only the vendor's interface address, authentication key, and timeout threshold need to be filled in to complete the integration. Configuration changes only affect newly established long-lived connections, ensuring that the store's ongoing fault handling process remains unaffected, guaranteeing uninterrupted store operations. Different nodes, such as fault text classification, handling plan generation, and customer complaint semantic extraction, send their unique scenario identifiers to the model scheduling module. The module then initiates a model selection algorithm to score the suitability of candidate models, selecting the optimal model across all dimensions to undertake the corresponding node's computation task. The module incorporates unified request conversion logic, automatically adapting to the differentiated interface formats of various vendors after receiving standardized upstream call parameters. After vendors return computation results, the data is uniformly formatted and sent back to the upstream, eliminating the development costs associated with inconsistent interface formats from multiple vendors. Each model call retains details such as the calling entity, model name, resource consumption, and execution time, and pushes this information synchronously to the tenant management module for persistent storage. This data is used for monthly statistics of total model usage and monitoring of quota consumption progress. When usage approaches the quota limit, the system automatically sends an alert to headquarters operations. Figure 3 As shown.

[0046] When offline store staff submit operational policy inquiries or report equipment / product malfunctions via the store's mobile chat mini-program, the intelligent routing module receives user dialogue requests in real time and retrieves all enabled intelligent agents within the tenant to perform matching and filtering. Headquarters administrators can independently switch matching execution strategies based on store traffic hours. During evening off-peak hours and complex customer complaint scenarios, a semantic priority strategy is activated, calling the intelligent matching algorithm to complete semantic matching calculations for all intelligent agents, adapting to business scenarios where the accuracy of responses meets preset thresholds, ensuring the accuracy of handling solutions. During peak hours and high-concurrency scenarios, a rapid matching strategy is switched, directly matching the corresponding intelligent agent based on pre-stored keyword-agent mapping relationships, eliminating the vectorization calculation step and shortening staff waiting time. After the intelligent routing module selects the store operation intelligent agent with the highest matching degree, it reads the agent's bound exclusive prompt words, associated knowledge base retrieval rules, and basic inference parameters. Using the unified computing power entry provided by the model scheduling module, it performs vector retrieval, concatenating the retrieved store operation standard knowledge base fragments and the original staff report content into a complete inference message, which is then sent to the large model to output basic consultation response content. If a store employee uploads a photo of a malfunctioning device and fills in the exception details to initiate a work order, the store exception handling business process identifier bound to the intelligent agent will trigger the complex business execution logic. The system will retrieve the complete definition of the published fault handling workflow, synchronize the dialogue context variables and retail enterprise tenant identity identifier to the process generation module, and drive the middle platform to generate a dedicated standardized workflow instance. During the instance's execution, the text classification and solution generation computing power requirements generated at each node continuously call the model scheduling module. After all nodes of the entire workflow have been executed, the system integrates the work order handling progress and basic Q&A content, generates a complete dialogue response including the person responsible for handling and the estimated processing time, and pushes it to the store employee's mobile app. The intelligent agent matching records, model resource consumption data, and workflow call records generated by the dialogue interaction are all synchronously archived to the tenant management module, facilitating monthly reviews of store fault handling efficiency by headquarters operations.

[0047] The capability open module exposes tenant-internal AI resources based on a standard long-connection protocol, while also supporting MCP standard protocol access. It provides SSE long-connection endpoints for standard MCP clients such as Cursor and Windsurf to access, and uses the JSON-RPC protocol for tool discovery and invocation. It supports access to external MCP servers via HTTP and STDIO transport protocols, automatically discovering and synchronizing the input schema and parameter definitions of external tools. The system has a built-in connector registry, registering four types of standard connectors: HTTP / REST connectors, MCP Server connectors, RPC connectors, and database query connectors. Each connector is pre-configured with authentication methods, permission boundaries, data anonymization strategies, and risk control rules. When the store's mobile app acts as an external client, it establishes a persistent long-connection session with a platform-issued exclusive identity credential. The module actively disconnects after the credential expires, mitigating the risk of unauthorized continuous external access. The store mini-program can initiate resource list query requests on demand. The module automatically scans all enabled agents, knowledge bases, tool skills, and published fault handling workflows within the tenant, encapsulating each type of resource into a standard tool description and transmitting it back via a long connection. When a store employee reports an equipment malfunction or queries replenishment specifications, the client initiates a resource call request. After identifying the resource type, the module distributes it to the process generation module, model scheduling module, or intelligent routing module for execution. Fault determination and handling notification results are transmitted back to the mini-program's front-end page in segments. Every external client call generates a complete audit log, recording the caller identifier, resource name, input / output summary, and overall execution time. This log is simultaneously pushed to the tenant management module for long-term retention, meeting the annual audit requirements of the enterprise's finance and risk control departments.

[0048] The tenant management module permeates all store business processes within the platform, implementing a four-tier organizational structure encompassing the platform, tenant companies, internal departments, and end users. All store work orders, intelligent agents, and knowledge base resources nationwide are uniquely linked to a retail enterprise tenant identifier. Every store reporting request carries tenant context information, preventing other retail brand tenants from accessing the operational data of this enterprise's stores and eliminating the risk of cross-tenant data leakage. The tenant management module includes multiple subscription versions, each bound to a corresponding set of available capabilities. Every time a fault handling workflow is invoked by a store's mini-program, the system automatically generates a complete operation record and single-node execution details, storing the input content, output content, execution status, and execution time of each node. All operational data is synchronously aggregated to the tenant management module's backend. Headquarters operations personnel can drag and drop to view the entire execution sequence through a visual interface, quickly identifying timeouts and solution matching failures, and adjusting node knowledge base retrieval thresholds and model call parameters accordingly. Headquarters operations personnel can only initiate operations through a dedicated natural language operations portal. After inputting natural language commands, the system's model scheduling module parses the command operation type and associated parameters. The scope of operations covers resource creation, quota adjustment, permission change, usage query, and service start / stop. The system automatically classifies operations into risk levels based on built-in risk level definitions. Level 1 low-risk operations, such as usage query and configuration viewing, are executed directly. Level 2 medium-risk operations, such as adding a store consultation agent, expanding knowledge base storage capacity, and increasing multimodal model call limits, take effect immediately. Level 3 high-risk operations, such as deleting historical fault handling workflows, reducing external tool concurrent quotas, and revoking regional operation approval permissions, require a secondary confirmation pop-up. The system will only execute the change after the operations personnel confirm that everything is correct. All operations are fully logged, recording the operator's account, operation time, operation content, and configuration status before and after the change. The risk control department can export the logs at any time to complete compliance checks.

[0049] This implementation of anomaly handling for chain retail stores fully leverages the system's end-to-end AI orchestration and multi-terminal integration capabilities. The tenant management module establishes a four-tier organizational structure to match the hierarchical management needs of stores nationwide, constraining the scale of store resource usage through quotas and permission checks. The process generation module automatically generates standardized fault handling workflows based on natural language requirements, supporting iterative releases and end-to-end operational traceability. The model scheduling module utilizes model selection algorithms to adapt to fault classifications and generate various scenario models, achieving unified scheduling of computing power from multiple vendors. The intelligent routing module flexibly switches matching strategies, using intelligent matching algorithms to balance consultation accuracy with peak concurrency latency. The capability openness module maintains a long-connection interface with the store's mobile mini-program. After implementation, the timeliness of store fault handling is optimized, and a complete data isolation, audit log, and hierarchical operation and maintenance risk management system are fully implemented, adapting to the high-concurrency retail business operation requirements of multiple stores.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An enterprise intelligent middleware system based on natural language and visual orchestration, characterized in that, The system includes: The workflow generation module receives natural language business requirements, parses them into node specification mappings, and generates a standardized directed acyclic graph (DAG) workflow topology after topology verification. It provides a visual drag-and-drop editing interface, supports workflow version management and runtime tracking, and outputs standardized workflow data. The natural language business requirements include at least one of six categories: business scenario objectives, data flow logic, knowledge base call requirements, model output format constraints, branch judgment conditions, and result output specifications. The workflow generation module retrieves built-in complete node specifications, hard constraints on the workflow topology, and structured prompt word templates. It then concatenates the requirement information with the templates and sends them to the model scheduling module for inference. After the model returns the original topology JSON data, the workflow generation module sequentially performs node type whitelist verification, topology island and loop detection, and start / end node integrity verification. It automatically corrects any verification anomalies and automatically arranges canvas node coordinates and completes default inference parameter values ​​after successful verification, generating standardized workflow data which is then pushed to the visual interface for rendering. The model scheduling module receives standardized workflow data, incorporates a model selection algorithm, uniformly stores all model metadata, automatically filters suitable models based on scene identifiers from upstream sources, uniformly encapsulates model interfaces from different vendors to provide a unified computing power access point, and records all usage data. The model selection algorithm is used for scene-model compatibility scoring. After receiving scene identifiers, it first filters a set of candidate models that meet basic capability requirements, and then performs a scoring operation on all models in the set. The operation expression is as follows: In the formula This represents the fit score of the m-th single candidate model for the current scene; The weight of the dimension representing the ability matching; Representative ability matching factor; Represents the weight of performance dimensions; Represents performance factor; This represents the weight of the cost control dimension; Represents cost factor; Intelligent routing module: Relying on a unified computing power entry point, with a built-in intelligent matching algorithm, it receives user dialogue requests, performs matching and filtering on all available intelligent agents within the tenant, selects the optimal intelligent agent to execute dialogue tasks, and calls the workflow generated by the process generation module to complete complex business. Capability Opening Module: Based on the standard long connection protocol and MCP standard protocol, it exposes the tenant's internal AI resources to the outside world, supports resource discovery and invocation by external clients, and links the process generation module, model scheduling module, and intelligent routing module to complete capability execution and generate full audit logs; Tenant Management Module: Covers all aspects of the process generation module, model scheduling module, intelligent routing module, and capability opening module. It supports a four-level organizational structure, manages tenants' ability to subscribe to AI resources on the platform and their fine-grained permissions, provides a natural language operation and maintenance entry point, issues permissions and quota rules to the process generation module, model scheduling module, intelligent routing module, and capability opening module, and aggregates the operational data of the entire platform.

2. The enterprise intelligent middleware system based on natural language and visual orchestration according to claim 1, characterized in that, The model scheduling module stores all model metadata. Each piece of metadata includes the model's vendor, model category, functional support status, vector dimension, and input / output modal attributes. It supports dynamically adding model vendor configurations; simply filling in the vendor's interface address, authentication key, and timeout threshold is sufficient for integration. Configuration changes take effect in real time without requiring a service restart. It has built-in unified request conversion logic. After receiving standardized call parameters from upstream, it automatically converts them into the request format required by the corresponding vendor. After obtaining the vendor's return results, it uniformly organizes them into a standard format and sends them back to upstream, completely shielding the interface differences between different vendors. Each call records the calling entity, model name, resource consumption, and execution time data, which is synchronized to the tenant management module for quota management and usage statistics.

3. The enterprise intelligent middleware system based on natural language and visual orchestration according to claim 1, characterized in that, The intelligent routing module includes two sets of matching execution strategies, and the tenant administrator can independently select the effective strategy according to the business scenario; The first strategy is a semantic-first approach, which calls the intelligent matching algorithm to perform semantic matching operations on all intelligent agents, adapting to business scenarios where the accuracy of the answers reaches a preset threshold. The second strategy is a fast matching approach, which, based on the pre-stored mapping relationship between keywords and intelligent agents, compares the keywords contained in the query text to directly match the corresponding intelligent agent, skipping the vectorization operation step, adapting to high-concurrency and low-latency business scenarios.

4. The enterprise intelligent middleware system based on natural language and visual orchestration according to claim 1, characterized in that, After selecting the optimal agent, the intelligent routing module executes the dialogue task. It first reads the agent's bound exclusive prompts, associated knowledge base retrieval rules, and basic inference parameters. Relying on a unified computing power entry point, it initiates a vector retrieval, concatenates the retrieved knowledge base fragments and the user's original dialogue request into a complete inference message, and sends it into the large model to complete the basic question-and-answer output. When the optimal agent is configured to bind a business process identifier, it triggers the composite business execution logic. Based on the business process identifier bound to the agent, it retrieves the corresponding published workflow definition, transmits the current dialogue context variables and tenant identity information to the workflow generation module, and drives the workflow generation module to start the corresponding standardized workflow instance. The computing power requirements of each node in the workflow continuously call the model scheduling module to support the computation. After the entire workflow is fully executed, the process output results are summarized and integrated with the basic question-and-answer content to generate the final dialogue response. The matching records, model consumption, and workflow call records generated throughout the dialogue process are synchronously pushed to the tenant management module for archiving.

5. The enterprise intelligent middleware system based on natural language and visual orchestration according to claim 1, characterized in that, The capability opening module provides a standard long connection call address to the outside world. When an external client initiates a connection request, it must carry a unique identity credential. After the credential is verified, a persistent long connection session is established. When the credential expires, the connection is actively disconnected. When a client initiates a resource list query request, it automatically scans all enabled agents, knowledge bases, skills, and published workflows within the current tenant, encapsulates each resource into a standard tool description, and returns it via a long connection. When a client initiates a resource call request, it parses the type of the called resource and distributes it to the process generation module, model scheduling module, or intelligent routing module to execute the corresponding operation. The execution result is returned in segments via a long connection. Each call generates an audit log, which includes the caller identifier, resource name, input / output summary, and execution time data, and is synchronized to the tenant management module for retention.

6. The enterprise intelligent middleware system based on natural language and visual orchestration according to claim 1, characterized in that, The tenant management module provides a dedicated natural language operation and maintenance entry point, which is only open to administrators with operation and maintenance permissions. After the administrators input natural language operation and maintenance commands, the linkage model scheduling module executes command parsing, identifies the operation type and corresponding parameters, and the operation types cover resource creation, quota adjustment, permission change, usage query, and service start and stop. After parsing, the corresponding internal interface is called to perform the actual operation. The operation risk is automatically determined according to the risk level definition. High-risk operations trigger a secondary confirmation process, and the operation is executed only after the confirmation is correct. After the operation is completed, the results will be compiled into natural language text and fed back to the administrators; all operation and maintenance operations will be logged, recording the operator, operation time, operation content, and status before and after the change, to meet the audit and traceability requirements.

Citation Information

Patent Citations

  • Multi-tenant resource dynamic access method based on MCP protocol, cloud hosting system and device

    CN121333769A

  • Enterprise-oriented AI service unified management method and system combined with MCP

    CN121585571A

  • Multi-agent dynamic workflow arrangement method and system based on MCP protocol

    CN122450655A