Multi-model integrated and managed ai agent system

Abstract

The application belongs to the technical field of artificial intelligence application, and discloses an AI intelligent agent system for multi-model integration and management, which comprises a model access unit, a knowledge processing unit, a process arrangement unit and an execution control unit. The model access unit calls a configuration template and generates a model call configuration according to a supplier type; the knowledge processing unit generates knowledge index data according to a knowledge base configuration; the process arrangement unit generates an intelligent agent process configuration comprising an execution node, a node input / output relationship and a resource reference relationship; and the execution control unit checks a model, a knowledge base, an external tool and a node input source before publishing, and completes knowledge retrieval, model request assembly, model request conversion and node output transmission during runtime. In the above manner, different supplier models, enterprise knowledge data and intelligent agent processes are collaboratively run in the same execution control chain.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence application technology, specifically an AI intelligent agent system for multi-model integration and management.

[0002] This invention can also be applied to scenarios such as grain storage and grain and oil processing engineering, central grain reserve storage management, engineering design project management, equipment operation and maintenance, enterprise knowledge Q&A, business process assistance, content generation and business system linkage, and is used to unify and manage the execution of various artificial intelligence models, enterprise knowledge data and intelligent agent processes. Background Technology

[0003] As enterprise business systems increasingly demand the use of artificial intelligence models, AI agents are gradually being applied to scenarios such as knowledge question answering, process assistance, content generation, and business system linkage. Existing agent systems typically integrate different types of models, such as large language models, vector models, rearrangement models, speech recognition models, or visual models, and combine them with enterprise knowledge base retrieval and process node orchestration to build AI applications for specific business needs.

[0004] In industry applications such as grain storage and processing engineering, central grain reserve storage management, engineering design project management, and equipment operation and maintenance, business data typically includes warehouse design data, process specifications, equipment operation records, quality inspection data, inventory management rules, project delivery documents, and operation and maintenance knowledge documents. These data sources are scattered, have significantly different formats, and often need to be integrated with intelligent agent processes such as question and answer generation, process approval, equipment inspection, report generation, and business system interface linkage. Therefore, this type of application scenario has high requirements for consistency between multi-model access, enterprise knowledge indexing, process node orchestration, and operation control.

[0005] In existing technologies, multi-model access typically involves configuring the calling address, authentication credentials, model parameters, and request format separately according to the interface rules of different vendors; knowledge augmentation generally involves segmenting, vectorizing, and indexing enterprise documents, web page data, or business materials for storage, and then returning knowledge content through vector retrieval, full-text retrieval, or hybrid retrieval; process orchestration generally organizes the execution order through the connection relationships between start nodes, model calling nodes, knowledge retrieval nodes, judgment nodes, loop nodes, and tool calling nodes. The above methods can complete model calling, knowledge question answering, and process execution to a certain extent, but model configuration, knowledge index data, process nodes, and external tools are usually configured and managed separately, lacking a unified reference relationship that runs through pre-release verification and runtime calls.

[0006] In actual operation, different vendor models have different authentication methods, request fields, and parameter formats. If a process node directly depends on the original interface of a certain vendor model, subsequent model replacement or parameter adjustment can easily affect the input and output configuration of the process node. Although the enterprise knowledge base can provide search content, if there is no clear correspondence between the knowledge base identifier, vector model identifier, knowledge index data, and process node, it is difficult to maintain consistency in the source, transmission path, and writing location of the knowledge content used when calling the model. For intelligent agent processes that include multi-node data transmission, conditional branches, loop processing, and external tool calls, if the consistency of model call configuration, knowledge index data, external tools, and node input data sources is not verified before release, problems such as inconsistent model request formats, incorrect transmission of knowledge content, unavailability of external tools, or missing node inputs may occur during runtime, thus affecting the continuous execution and reproducible operation of the intelligent agent process. Summary of the Invention

[0007] The purpose of this invention is to provide an AI intelligent agent system for multi-model integration and management, so as to solve the problems mentioned in the background art.

[0008] To address the problems existing in enterprises when building artificial intelligence applications, such as inconsistent access rules for different vendor models, difficulty in reliably incorporating enterprise knowledge into model calls, and lack of unified operational control for complex business processes and model capabilities, this invention provides an AI intelligent agent system for multi-model integration and management; the system includes a model access unit, a knowledge processing unit, a process orchestration unit, and an execution control unit.

[0009] The model access unit is used to receive model access parameters from artificial intelligence models. The model access parameters include supplier type, model name, model type, basic model identifier, calling address, authentication credentials, and extended parameters. The model access unit calls the corresponding configuration template according to the supplier type and generates model calling configuration according to the configuration template and model access parameters. The model calling configuration includes model type, calling address, authentication rules, unified calling fields, and parameter conversion rules. The configuration template includes general fields and supplier-specific fields. The general fields are used to record model access parameters shared by different artificial intelligence models, and the supplier-specific fields are used to record authentication parameters, model parameters, and calling format parameters corresponding to different supplier models.

[0010] Therefore, after different vendor models are integrated, they all participate in the subsequent process operation in the form of model call configuration, and the execution node does not need to record the original call rules of each vendor model separately.

[0011] The model access unit includes a template parsing module and a configuration generation module. The template parsing module determines the general fields and supplier-specific fields that need to be entered and verified based on the supplier type. The configuration generation module writes the parameters in the general fields into the unified call fields and writes the parameters in the supplier-specific fields into the authentication rules or parameter conversion rules. Thus, the configuration template is not only used to receive model access parameters, but also to form the data foundation required for subsequent model call request conversion, enabling the execution control unit to assemble the internal system call request based on the unified call fields and convert it into the model call request required by the corresponding artificial intelligence model according to the parameter conversion rules.

[0012] The knowledge processing unit is used to receive business knowledge data; the sources of business knowledge data include offline documents, web page data, and process execution data; the knowledge processing unit determines the vector model call configuration from the model call configuration according to the knowledge base identifier, segmentation rules, and vector model identifier in the knowledge base configuration, and calls the vector model corresponding to the vector model call configuration to perform vectorization processing on the knowledge segments formed by the business knowledge data, generating knowledge index data associated with the knowledge base identifier; the knowledge index data includes the knowledge base identifier, knowledge segment text, knowledge segment vector, knowledge tag, and vector model identifier.

[0013] Therefore, before business knowledge data is invoked by the intelligent agent process, it is first formed into knowledge index data with knowledge base identifier and vector model identifier, which facilitates subsequent knowledge retrieval nodes to retrieve data according to the determined knowledge base scope.

[0014] The knowledge processing unit includes a knowledge import module, a segmentation processing module, a vectorization module, an index storage module, and a retrieval processing module; the knowledge import module is used to receive offline documents, web page data, and process execution data; the segmentation processing module generates knowledge segments according to document title level, segment identifier, regular expression, or segment character count threshold.

[0015] The vectorization module calls the vector model to generate knowledge segment vectors; the index storage module associates and stores the knowledge base identifier, knowledge segment text, knowledge segment vectors, knowledge tags, and vector model identifiers as knowledge index data; the retrieval processing module performs vector retrieval, full-text retrieval, or hybrid retrieval on the knowledge index data based on the knowledge base identifier referenced by the knowledge retrieval node; when performing vector retrieval or hybrid retrieval, the retrieval processing module filters the retrieval results based on a similarity threshold; when performing full-text retrieval, the retrieval processing module forms retrieval results based on keyword matching results; the retrieval processing module determines the knowledge content output by the knowledge retrieval node from the retrieval results according to the number of referenced segments and the maximum number of referenced characters; thus, the formation process of knowledge content has a defined data source, retrieval scope, and output rules.

[0016] When the vector model identifier corresponding to the target knowledge base changes, the knowledge processing unit reads the knowledge segment text already stored in the target knowledge base, calls the changed vector model to regenerate the knowledge segment vector, and stores the regenerated knowledge segment vector as associated with the changed vector model identifier. Thus, the knowledge segment vector in the knowledge index data maintains a correspondence with the currently used vector model identifier, avoiding subsequent knowledge retrieval from still calling the knowledge segment vector generated by the original vector model.

[0017] The process orchestration unit is used to generate intelligent agent process configuration based on business process configuration; the intelligent agent process configuration includes multiple execution nodes, node connection relationships, node input-output relationships, node parameters, and resource reference relationships of each execution node; the execution node includes a model call node; when the resource reference relationship includes a knowledge base identifier, the execution node includes a knowledge retrieval node; the execution node also includes one or more of the following: start node, judgment node, loop node, and tool call node.

[0018] Resource reference relationships are used to represent the model call configuration, knowledge base identifier, or external tool referenced by the corresponding execution node; node input-output relationships include variable mapping relationships, which are used to make the output data of the previous execution node the input data of the next execution node; thus, the agent process configuration not only records the connection order of execution nodes, but also records the resources that each execution node needs to call and the data transfer relationships between execution nodes.

[0019] The execution control unit is used to verify, before the intelligent agent process configuration is released, whether the model call configuration referenced by each execution node exists, whether the referenced knowledge base identifier corresponds to the generated knowledge index data, and whether the referenced external tools exist, based on the resource reference relationship; at the same time, it verifies, based on the node input-output relationship, whether the input data of each execution node is provided by user input data, the output data of the previous execution node, or node parameters.

[0020] After verification, the execution control unit generates an executable process configuration; thus, before the intelligent agent process is released, it can be confirmed that the model, knowledge base, external tools, and node input data sources all have the conditions for execution, avoiding processes that lack running resources or input sources from entering the callable state.

[0021] The execution control unit is also used to determine the current execution node according to the executable flow configuration after receiving user input data, and generate a call request for the current execution node based on the resource reference relationship, node input-output relationship and node parameters of the current execution node; when the current execution node is a knowledge retrieval node, the execution control unit retrieves the knowledge content from the knowledge index data according to the knowledge base identifier referenced by the current execution node, and uses the knowledge content as the output data of the knowledge retrieval node.

[0022] When the current execution node is a model invocation node, the execution control unit assembles the input data and node parameters of the current execution node into an internal system invocation request based on the unified invocation field in the model invocation configuration referenced by the current execution node. When the output data of the knowledge retrieval node, which records the node input-output relationship, is mapped to the input data of the current execution node, the execution control unit writes the knowledge content output by the knowledge retrieval node into the internal system invocation request. Subsequently, the execution control unit converts the internal system invocation request into a model invocation request required by the artificial intelligence model referenced by the current execution node according to the parameter conversion rules in the model invocation configuration referenced by the current execution node, and receives the output data of the current execution node returned by the artificial intelligence model, and passes the output data to subsequent execution nodes. Thus, the model invocation node can receive user input data, node parameters, and knowledge content under the same execution control logic, and complete the conversion of invocation requests to different vendor models through the model invocation configuration.

[0023] When an execution node includes a tool invocation node, the external tool referenced by the tool invocation node has tool startup parameters and tool input parameters, which are stored separately. When executing a tool invocation node, the execution control unit reads the tool startup parameters according to the resource reference relationship of the tool invocation node, generates tool input parameters based on the output data of the previous execution node, and passes the tool execution result as the output data of the tool invocation node to subsequent execution nodes.

[0024] Therefore, the fixed configuration parameters of external tools are handled separately from the runtime input parameters, avoiding the mixing of tool configuration parameters and execution node data.

[0025] The execution control unit is also used to generate an agent application interface after the executable process configuration is published, and to restrict the model call configuration, knowledge base identifier and external tools that the agent application interface can call when executing the executable process configuration according to the resource permissions corresponding to the user identity.

[0026] Therefore, when external systems call executable process configurations through the intelligent agent application interface, they still use models, knowledge bases, and external tools according to resource permissions, ensuring that the process calls after release are consistent with the internal resource authorization relationships of the system.

[0027] The beneficial effects of this invention are as follows: 1. This invention uses a model access unit to call a configuration template based on the supplier type, and generates a model call configuration from the configuration template and model access parameters. This ensures that the model type, call address, authentication rules, unified call fields, and parameter conversion rules of different supplier AI models are uniformly recorded during the access phase. When the execution control unit runs the model call node, it first assembles the internal system call request based on the unified call fields, and then converts it into the model call request required by the current AI model based on the parameter conversion rules. Through the above processing, the process nodes no longer directly depend on the original interface rules of each supplier model. A unified connection is formed between the model access configuration and the process execution call, thereby reducing the problem of repeated adaptation caused by differences in supplier interfaces, authentication methods, and parameter formats. This allows adjustments to be made at the model call configuration layer when replacing models or switching between similar models. 2. This invention uses a knowledge processing unit to segment, vectorize, and index business knowledge data based on knowledge base identifiers, segmentation rules, and vector model identifiers, forming knowledge index data that includes knowledge base identifiers, knowledge segment text, knowledge segment vectors, knowledge tags, and vector model identifiers. The process orchestration unit configures the knowledge base identifiers to knowledge retrieval nodes through resource reference relationships. During runtime, the execution control unit retrieves knowledge content from the knowledge index data based on the knowledge base identifiers and writes the knowledge content into the system's internal call request for the model call node through node input-output relationships. Through this processing, enterprise documents, web page data, and process execution data can participate in model calls with a clear knowledge base scope, retrieval rules, and node transmission relationships, avoiding the disordered splicing of knowledge content into model requests as temporary text. This ensures that the source, transmission path, and usage location of the knowledge content can be determined in the agent's process configuration. 3. This invention generates an intelligent agent process configuration that includes execution nodes, node connection relationships, node input-output relationships, node parameters, and resource reference relationships through a process orchestration unit. Before release, the execution control unit verifies whether the model call configuration exists, whether the knowledge base identifier corresponds to the generated knowledge index data, whether external tools exist, and whether the source of node input data is complete. Only after the verification is passed can an executable process configuration be generated. During the operation phase, the execution control unit completes knowledge retrieval, model request assembly, model request conversion, tool invocation, and node output transmission in sequence according to the executable process configuration. Through the above processing, model access, knowledge processing, and process orchestration are no longer independent functional modules, but form a continuous control chain in the two stages of release verification and operation execution, thereby reducing the problem of process interruption caused by resource lack, input source lack, or inconsistent model request format. Attached Figure Description

[0028] Figure 1This is a schematic diagram illustrating the overall architecture and operation flow of the AI ​​intelligent agent system for multi-model integration and management according to the present invention. Figure 2 This is a flowchart illustrating the overall execution process of the AI ​​intelligent agent system for multi-model integration and management according to the present invention. Figure 3 This is a flowchart illustrating the model access and configuration generation process of this invention. Figure 4 This is a flowchart illustrating the runtime process of the control unit of this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figures 1 to 4 As shown, this embodiment focuses on the data flow relationship between model access, knowledge processing, process orchestration and execution control, enabling those skilled in the art to complete system deployment and operation configuration based on this embodiment.

[0031] The AI ​​intelligent agent system for multi-model integration and management provided in this embodiment can be used in, but is not limited to, technical fields and application scenarios such as grain storage and grain and oil processing engineering, central grain reserve storage management, engineering design project management, equipment operation and maintenance, enterprise knowledge Q&A, business process assistance, content generation, and business system linkage. For example, in the central grain reserve storage management scenario related to China Grain Reserves Corporation (Sinograin), business knowledge data may include grain storage system documents, warehouse design data, grain and oil processing technology data, equipment operation and maintenance data, quality inspection records, inventory management rules, inspection records, report templates, and engineering project data. After the above business knowledge data is used by the knowledge processing unit to generate knowledge index data, it can be configured by the process orchestration unit as a resource reference object for knowledge retrieval nodes, model call nodes, judgment nodes, or tool call nodes, and called by the execution control unit in the process of Q&A generation, inspection record summarization, quality data query, inventory rule query, equipment operation and maintenance assistance, report generation, and business system interface linkage.

[0032] The AI ​​agent system for multi-model integration and management provided in this embodiment includes a model access unit, a knowledge processing unit, a process orchestration unit, and an execution control unit. The model access unit generates model call configurations; the knowledge processing unit generates knowledge index data; the process orchestration unit generates agent process configurations; and the execution control unit performs pre-release verification of the agent process configurations and completes knowledge retrieval, model invocation, tool invocation, and node output transmission according to the executable process configurations during runtime.

[0033] The model access unit receives model access parameters from the artificial intelligence model. These parameters include vendor type, model name, model type, basic model identifier, calling address, authentication credentials, and extended parameters. The vendor type is used to determine the configuration template; the model name is used to distinguish different artificial intelligence models within the system; the model type is used to determine the calling category of the artificial intelligence model in the process; the basic model identifier is used to record the model identifier of the called model on the vendor side or in the local deployment environment; the calling address is used to record the model service access address; the authentication credentials are used to generate authentication rules; and the extended parameters are used to record the additional model parameters or calling format parameters required by the vendor model when calling it.

[0034] The model types include two or more of the following: large language model, vector model, reordering model, speech recognition model, speech synthesis model, visual model, image generation model, and video generation model. Large language models are used for text generation or question-answering generation; vector models are used to generate knowledge segmentation vectors; reordering models are used to rank candidate retrieval results; speech recognition models are used to convert speech data into text data; speech synthesis models are used to convert text data into speech data; visual models are used for image recognition or image understanding; image generation models are used to generate images based on node input data; and video generation models are used to generate videos based on node input data.

[0035] The model access unit calls the corresponding configuration template based on the supplier type. The configuration template includes general fields and supplier-specific fields. The general fields include the model name field, model type field, basic model identifier field, and call address field. The supplier-specific fields include one or more of the authentication parameter field, model parameter field, and call format parameter field. The authentication parameter field is used to generate authentication rules. The model parameter field is used to record the parameters required for model runtime. The call format parameter field is used to generate parameter conversion rules.

[0036] The model access unit includes a template parsing module and a configuration generation module. The template parsing module determines the general fields and supplier-specific fields that need to be entered and verified based on the supplier type. The configuration generation module generates the model call configuration based on the general fields and supplier-specific fields. Specifically, the configuration generation module writes the parameters from the general fields into the unified call field, the authentication parameters from the supplier-specific fields into the authentication rules, and the model parameters and call format parameters from the supplier-specific fields into the parameter conversion rules. The generated model call configuration includes the model type, call address, authentication rules, unified call field, and parameter conversion rules.

[0037] The unified invocation fields are the set of fields used by the system to assemble model invocation requests. They are used to receive the input data, node parameters, context variables, and knowledge content of the current execution node. The authentication rules record the writing position and writing method of authentication credentials in the model invocation request. The parameter conversion rules record the mapping relationship between the fields in the system's internal invocation request and the fields required by the current artificial intelligence model. The execution control unit forms the system's internal invocation request based on the unified invocation fields and converts the system's internal invocation request into the model invocation request required by the current artificial intelligence model according to the parameter conversion rules.

[0038] When multiple model call configurations correspond to the same model type, the execution control unit determines the model call configuration referenced by the current execution node based on the model's activation status, call priority, and model resource permissions.

[0039] Model enable status indicates whether the model call configuration is allowed to be invoked; call priority is used to determine the selection order among multiple callable model call configurations; model resource permissions indicate whether the current user or the current process has permission to invoke the model call configuration.

[0040] The knowledge processing unit receives business knowledge data; business knowledge data includes one or more of the following: offline documents, web page data, and process execution data.

[0041] Offline documents include internal enterprise documents, instructional documents, Q&A materials, or business specification documents; web page data includes the text content of pages obtained through web page addresses; process operation data includes data generated during the operation of intelligent agent processes and allowed to be written into the knowledge base.

[0042] After business knowledge data enters the knowledge processing unit, it undergoes segmentation, vectorization, and index storage to form knowledge index data associated with the knowledge base identifier.

[0043] The knowledge processing unit processes business knowledge data according to the knowledge base configuration; the knowledge base configuration includes a knowledge base identifier, segmentation rules, and a vector model identifier; the knowledge base identifier is used to determine the knowledge base scope to which the business knowledge data belongs; the segmentation rules are used to determine the segmentation method of the business knowledge data; and the vector model identifier is used to determine the vector model used to generate knowledge segmentation vectors.

[0044] The knowledge processing unit determines the vector model invocation configuration from the model invocation configuration generated by the model access unit based on the vector model identifier, and calls the vector model corresponding to the vector model invocation configuration to perform vectorization processing on the knowledge segments formed by the business knowledge data.

[0045] The knowledge processing unit includes a knowledge import module, a segmentation module, a vectorization module, an index storage module, and a retrieval processing module. The knowledge import module receives offline documents, web page data, and process execution data. The segmentation module generates knowledge segments according to document title level, segment identifier, regular expression, or segment character count threshold.

[0046] The document heading level is used to determine the segment boundaries based on the heading structure in the document; the segment identifier is used to determine the segment boundaries based on preset characters or paragraph marks; the regular expression is used to identify the segment boundaries based on preset matching rules; and the segment character count threshold is used to limit the character length of a single knowledge segment.

[0047] The vectorization module calls the vector model to generate knowledge segment vectors; the index storage module associates and stores the knowledge base identifier, knowledge segment text, knowledge segment vector, knowledge tag, and vector model identifier as knowledge index data.

[0048] The knowledge index data includes knowledge base identifiers, knowledge segment texts, knowledge segment vectors, knowledge tags, and vector model identifiers. The knowledge base identifier is used to determine the knowledge base to which the knowledge segment belongs. The knowledge segment text is used to store the text content that can be retrieved and returned. The knowledge segment vector is used to participate in vector retrieval. The knowledge tag is used to classify or filter the knowledge segments. The vector model identifier is used to record the vector model used to generate the knowledge segment vector.

[0049] Through the aforementioned associated storage, knowledge retrieval nodes can perform searches within a defined knowledge index data range based on the knowledge base identifier.

[0050] The retrieval processing module performs vector retrieval, full-text retrieval, or hybrid retrieval on the knowledge index data. When performing vector retrieval, the retrieval processing module generates retrieval text based on the input data of the knowledge retrieval node, calls the vector model in the knowledge base configuration to generate retrieval vectors, compares the similarity of the retrieval vectors with the knowledge segment vectors, and filters the retrieval results based on the similarity threshold. When performing full-text retrieval, the retrieval processing module extracts keywords based on the input data of the knowledge retrieval node and forms retrieval results based on the keyword matching results.

[0051] When performing a hybrid search, the search processing module generates vector search results and full-text search results respectively. The vector search results are filtered according to a similarity threshold, and the filtered vector search results are merged with the full-text search results to form the search results.

[0052] The retrieval processing module determines the knowledge content output by the knowledge retrieval node from the retrieval results based on the number of reference segments and the maximum number of reference characters.

[0053] When the vector model identifier corresponding to the target knowledge base changes, the knowledge processing unit reads the knowledge segment text already stored in the target knowledge base, calls the changed vector model to regenerate the knowledge segment vector, and associates and stores the regenerated knowledge segment vector with the changed vector model identifier.

[0054] After regeneration, subsequent retrievals of the target knowledge base will use the modified vector model to identify the corresponding knowledge segment vectors.

[0055] The process orchestration unit generates intelligent agent process configuration based on the business process configuration; the business process configuration is configured by the user in the process orchestration interface, or generated by the system based on a saved process template; the intelligent agent process configuration includes multiple execution nodes, node connection relationships, node input-output relationships, node parameters, and resource reference relationships for each execution node; execution nodes include model call nodes; when the resource reference relationship includes a knowledge base identifier, the execution node includes a knowledge retrieval node; the execution node also includes one or more of the following: start node, judgment node, loop node, and tool call node.

[0056] Node connection relationships are used to represent the sequential connection between execution nodes; node input-output relationships are used to represent how the output data of the previous execution node is used as the input data of the next execution node; node parameters are used to store the fixed configuration content of the corresponding execution node.

[0057] The node parameters for model invocation nodes include prompt text, output format requirements, or model running parameters; the node parameters for knowledge retrieval nodes include retrieval method, similarity threshold, number of reference segments, or maximum number of reference characters; the node parameters for judgment nodes include conditional expressions; the node parameters for loop nodes include loop object, number of loops, or loop interruption conditions; the node parameters for tool invocation nodes include external tool identifiers and input parameter mapping rules; resource reference relationships are used to represent the model invocation configuration, knowledge base identifier, or external tool referenced by the corresponding execution node.

[0058] The node input-output relationship includes variable mapping relationships; variable mapping relationships are used to make the output data of the previous execution node the input data of the next execution node; the knowledge content output by the knowledge retrieval node can be written into the input data of the model calling node through variable mapping relationships; the tool execution results output by the tool calling node can be written into the input data of the subsequent model calling node or judgment node through variable mapping relationships; through the node input-output relationship, the agent process configuration can simultaneously record the execution order and data transmission method between execution nodes.

[0059] The execution control unit performs verification before the intelligent agent process configuration is released; the execution control unit verifies whether the model call configuration referenced by each execution node exists according to the resource reference relationship, verifies whether the knowledge base identifier referenced by each execution node corresponds to the generated knowledge index data, and verifies whether the external tools referenced by each execution node exist.

[0060] The execution control unit also verifies whether the input data of each execution node is provided by user input data, output data of the previous execution node, or node parameters based on the node input-output relationship. After the verification is passed, the execution control unit generates an executable process configuration. The executable process configuration includes the verified execution nodes, node connection relationships, node input-output relationships, node parameters, and resource reference relationships.

[0061] If the model call configuration referenced by the model call node does not exist, the execution control unit will not generate an executable process configuration; if the knowledge base identifier referenced by the knowledge retrieval node does not correspond to the generated knowledge index data, the execution control unit will not generate an executable process configuration; if the external tool referenced by the tool call node does not exist, the execution control unit will not generate an executable process configuration; if the input data of any execution node cannot be provided by user input data, the output data of the previous execution node, or node parameters, the execution control unit will not generate an executable process configuration; through the above verification, the agent process enters the callable state after the resource reference relationship and node input-output relationship meet the execution conditions.

[0062] After receiving user input data, the execution control unit determines the current execution node according to the executable flow configuration, and generates a call request for the current execution node based on the resource reference relationship, node input-output relationship and node parameters of the current execution node.

[0063] When the current execution node is the start node, the execution control unit uses the user input data as the output data of the start node or the input data of the subsequent execution nodes; when the current execution node is the knowledge retrieval node, the execution control unit retrieves the knowledge content from the knowledge index data according to the knowledge base identifier referenced by the current execution node, and uses the knowledge content as the output data of the knowledge retrieval node.

[0064] When the current execution node is a model calling node, the execution control unit assembles the input data and node parameters of the current execution node into an internal system calling request based on the unified calling field in the model calling configuration referenced by the current execution node. When the output data of the knowledge retrieval node in the node input-output relationship record is mapped to the input data of the current execution node, the execution control unit writes the knowledge content output by the knowledge retrieval node into the internal system calling request.

[0065] Subsequently, the execution control unit converts the internal system call request into a model call request required by the artificial intelligence model referenced by the current execution node according to the parameter conversion rules in the model call configuration of the current execution node. After the model call request is formed, the execution control unit initiates a call to the artificial intelligence model according to the call address and authentication rules in the model call configuration, and receives the output data of the current execution node returned by the artificial intelligence model, and passes the output data to the subsequent execution nodes.

[0066] An internal system call request consists of at least the input data and node parameters of the currently executing node; when the node input-output relationship maps the knowledge content output by the knowledge retrieval node to the input data of the currently executing node, the internal system call request also includes the knowledge content.

[0067] The field names in the internal system call requests are represented by a unified call field; the parameter conversion rules are used to write the data corresponding to the unified call field into the request field required by the current artificial intelligence model, so that different vendor models can complete the call conversion through the same execution control unit.

[0068] When the current execution node is a decision node, the execution control unit determines the subsequent execution node based on the conditional expression and input data of the decision node. When the current execution node is a loop node, the execution control unit determines whether to continue executing the nodes within the loop based on the loop object, the number of loop iterations, or the loop interruption condition. When the current execution node is a tool invocation node, the execution control unit reads the tool startup parameters of the external tool based on the resource reference relationship of the tool invocation node, and generates tool input parameters based on the output data of the previous execution node. The tool startup parameters and tool input parameters are then used together to generate a tool invocation request.

[0069] After the tool completes its execution, the execution control unit will use the tool execution result as the output data of the tool calling node and pass it to the subsequent execution node.

[0070] Tool startup parameters and tool input parameters are stored separately; tool startup parameters are parameters pre-configured by the external tool before execution, while tool input parameters are parameters generated based on the output data of the previous execution node during process execution; when the execution control unit generates a tool call request, it reads the tool startup parameters and tool input parameters respectively, and combines the two into a tool call request.

[0071] After the executable process configuration is published, the execution control unit generates the agent application interface; after the external system submits user input data through the agent application interface, the execution control unit executes the corresponding process according to the executable process configuration.

[0072] The execution control unit restricts the model call configuration, knowledge base identifier, and external tools that the intelligent agent application interface can call when executing executable process configurations, based on the resource permissions corresponding to the user's identity.

[0073] If the user does not have the corresponding resource permissions, the execution control unit will not allow the corresponding model to call configuration, knowledge base identifier or external tools to participate in the execution of this process.

[0074] In this embodiment, a continuous data processing relationship is formed between the model access unit, the knowledge processing unit, the process orchestration unit, and the execution control unit.

[0075] The model call configuration generated by the model access unit is used by the knowledge processing unit to determine the vector model call configuration, and is referenced by the process orchestration unit through resource reference relationships. It is also used by the execution control unit for model call request transformation.

[0076] The knowledge index data generated by the knowledge processing unit is referenced by the process orchestration unit through the knowledge base identifier, and is retrieved by the knowledge retrieval node at runtime to form knowledge content.

[0077] The intelligent agent process configuration generated by the process orchestration unit is verified by the execution control unit to form an executable process configuration. During runtime, the execution control unit completes node execution, data transmission, and resource invocation based on the executable process configuration.

[0078] Through the above processing, the model call configuration, knowledge index data, and intelligent agent process configuration form a continuous call relationship during the release verification phase and the runtime execution phase.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI intelligent agent system for multi-model integration and management, characterized in that: It includes a model access unit, a knowledge processing unit, a process orchestration unit, and an execution control unit; The model access unit is used to receive model access parameters of artificial intelligence models, call the corresponding configuration template according to the supplier type in the model access parameters, and generate model call configuration according to the configuration template and the model access parameters. The model call configuration includes model type, call address, authentication rules, unified call fields and parameter conversion rules. The knowledge processing unit is used to receive business knowledge data, determine the vector model call configuration from the model call configuration according to the knowledge base identifier, segmentation rules and vector model identifier in the knowledge base configuration, and call the vector model corresponding to the vector model call configuration to perform vectorization processing on the knowledge segments formed by the business knowledge data, and generate knowledge index data associated with the knowledge base identifier. The process orchestration unit is used to generate intelligent agent process configuration according to business process configuration. The intelligent agent process configuration includes multiple execution nodes, node connection relationships, node input and output relationships, node parameters, and resource reference relationships of each execution node. The multiple execution nodes include model calling nodes, and when the resource reference relationship includes a knowledge base identifier, the multiple execution nodes include a knowledge retrieval node. The resource reference relationship is used to indicate the model calling configuration, knowledge base identifier, or external tool referenced by the corresponding execution node. The execution control unit is used to verify, before the intelligent agent process configuration is released, whether the model call configuration referenced by each execution node exists, whether the referenced knowledge base identifier corresponds to the generated knowledge index data, and whether the referenced external tools exist, according to the resource reference relationship. It also verifies whether the input data of each execution node is provided by user input data, the output data of the previous execution node, or node parameters according to the node input-output relationship. After the verification is passed, an executable process configuration is generated. The execution control unit is also used to determine the current execution node according to the executable flow configuration after receiving user input data, and generate a call request for the current execution node based on the resource reference relationship, node input-output relationship and node parameters of the current execution node; When the current execution node is a knowledge retrieval node, the execution control unit retrieves knowledge content from the knowledge index data based on the knowledge base identifier referenced by the current execution node, and uses the knowledge content as the output data of the knowledge retrieval node; When the current execution node is a model calling node, the execution control unit assembles the input data and node parameters of the current execution node into an internal system calling request based on the unified calling field in the model calling configuration referenced by the current execution node. When the output data of the knowledge retrieval node recorded in the node input-output relationship is mapped to the input data of the current execution node, the execution control unit writes the knowledge content output by the knowledge retrieval node into the system internal call request; The execution control unit converts the internal system call request into a model call request required by the artificial intelligence model referenced by the current execution node according to the parameter conversion rules in the model call configuration of the current execution node, and receives the output data of the current execution node returned by the artificial intelligence model, and transmits the output data to the subsequent execution nodes.

2. The AI ​​intelligent agent system for multi-model integration and management according to claim 1, characterized in that: The model access parameters also include model name, model type, basic model identifier, calling address, authentication credentials, and extended parameters; the configuration template includes general fields and supplier-specific fields. The general fields are used to record model access parameters shared by different artificial intelligence models, and the supplier-specific fields are used to record one or more of the authentication parameters, model parameters, and calling format parameters corresponding to different supplier models.

3. The AI ​​intelligent agent system for multi-model integration and management according to claim 2, characterized in that: The model access unit includes a template parsing module and a configuration generation module; the template parsing module is used to determine the general fields and supplier-specific fields that need to be entered and verified according to the supplier type; the configuration generation module is used to write the parameters in the general fields into the unified call field, and write the parameters in the supplier-specific fields into the authentication rules or parameter conversion rules.

4. The AI ​​agent system for multi-model integration and management according to claim 3, characterized in that: The model type includes two or more of the following model types: large language model, vector model, rearrangement model, speech recognition model, speech synthesis model, visual model, image generation model, and video generation model; when the same model type corresponds to multiple model call configurations, the execution control unit determines the model call configuration referenced by the current execution node based on the model activation status, call priority, and model resource permissions.

5. The AI ​​agent system for multi-model integration and management according to claim 4, characterized in that: The knowledge processing unit includes a knowledge import module, a segmentation processing module, a vectorization module, and an index storage module; the knowledge import module is used to receive one or more types of business knowledge data from offline documents, web page data, and process execution data; the segmentation processing module is used to generate knowledge segments according to document title level, segmentation identifier, regular expression, or segment character count threshold; The vectorization module is used to call the vector model to generate knowledge segment vectors; the index storage module is used to associate and store the knowledge base identifier, knowledge segment text, knowledge segment vector, knowledge tag and vector model identifier as the knowledge index data.

6. The AI ​​agent system for multi-model integration and management according to claim 5, characterized in that: The knowledge processing unit further includes a retrieval processing module; the retrieval processing module is used to perform vector retrieval, full-text retrieval, or mixed retrieval on the knowledge index data; when performing vector retrieval or mixed retrieval, the retrieval processing module filters the retrieval results according to a similarity threshold; when performing full-text retrieval, the retrieval processing module forms retrieval results based on keyword matching results; the retrieval processing module determines the knowledge content output by the knowledge retrieval node from the retrieval results according to the number of cited segments and the maximum number of cited characters.

7. The AI ​​agent system for multi-model integration and management according to claim 6, characterized in that: The knowledge processing unit is used to read the knowledge segment text stored in the target knowledge base when the vector model identifier corresponding to the target knowledge base changes, call the changed vector model to regenerate the knowledge segment vector, and associate and store the regenerated knowledge segment vector with the changed vector model identifier.

8. The AI ​​agent system for multi-model integration and management according to claim 7, characterized in that: The plurality of execution nodes also includes one or more of start nodes, decision nodes, loop nodes, and tool call nodes; the node input-output relationship includes a variable mapping relationship, which is used to make the output data of the previous execution node the input data of the next execution node; The execution control unit determines the execution order of each execution node based on the node connection relationship, the conditional branch of the judgment node, or the loop condition of the loop node.

9. The AI ​​agent system for multi-model integration and management according to claim 8, characterized in that: The external tool referenced by the tool invocation node has tool startup parameters and tool input parameters, which are stored separately. When the execution control unit executes the tool invocation node, it reads the tool startup parameters according to the resource reference relationship of the tool invocation node, generates the tool input parameters according to the output data of the previous execution node, and uses the tool execution result as the output data of the tool invocation node and passes it to the subsequent execution node.

10. The AI ​​agent system for multi-model integration and management according to claim 9, characterized in that: The execution control unit is used to generate an agent application interface after the executable process configuration is published, and to restrict the model call configuration, knowledge base identifier and external tools that the agent application interface can call when executing the executable process configuration according to the resource permissions corresponding to the user identity.

Images