Intelligent workflow interaction configuration method and system based on large model

Abstract

The application discloses an intelligent workflow interaction configuration method and system based on a large model, belongs to the technical field of workflow configuration, and specifically comprises the following steps: obtaining function implementation requirement information of a user, combining a plurality of preset models, a knowledge base and plug-ins, and constructing an overall large model. The large model can output corresponding output requirement information based on input requirement information, and supports knowledge retrieval and configuration. In the construction process, steps such as model selection, knowledge base screening, plug-in integration and parameter adjustment are involved. The output process comprises generation, evaluation, decision analysis and optimized selection of a workflow configuration scheme. The application realizes the intelligentization and automation of workflow configuration, and improves the configuration efficiency and accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of workflow configuration, and in particular to an intelligent workflow interaction configuration method and system based on a large model. BACKGROUND

[0002] With the continuous development of enterprise business and the promotion of digital transformation, workflow management systems play a crucial role in enterprise operations. Traditional workflow configuration methods often require professional technical personnel to define and set through complex programming or the use of specific workflow modeling tools, which not only requires configuration personnel to have deep technical knowledge, but also makes the entire configuration process cumbersome and time-consuming, making it difficult to quickly respond to changes in business requirements. In addition, for non-technical business personnel, it is almost impossible to understand and participate in the configuration of the workflow, resulting in a communication gap between business and technology, further affecting the optimization and adjustment efficiency of the workflow.

[0003] A Chinese patent with authorization announcement number CN112148416B discloses an information interaction method, device, equipment and storage medium, comprising: when a mobile interaction device and a target scene device establish a communication connection, determining interaction authentication information corresponding to the target scene device; based on the interaction authentication information, obtaining interaction workflow information corresponding to the target scene device, and displaying the interaction workflow information on the mobile interaction device display interface. The technical solution of this embodiment can flexibly configure the interaction relationship between the mobile interaction device and the target scene device, and is especially suitable for the case of interacting with multiple target scene devices through one mobile interaction device. Moreover, based on the interaction authentication information, the interaction workflow information corresponding to the target scene device is obtained and displayed, which can facilitate users to view and operate the interaction workflow information, and realize the interaction between the mobile interaction device and the target scene device in a targeted manner.

[0004] A Chinese patent with authorization announcement number CN117406979B discloses an interface interaction design method and system for computing workflow, comprising: workflow orchestration: creating a new workflow and filling in basic configuration information in the computing workflow list interface, selecting and connecting different types of computing units in the workflow canvas interface to obtain a computing workflow, and configuring and submitting the computing of each computing unit in the form of a drawer floating layer interaction; workflow execution: viewing the running situation of the computing workflow in the workflow canvas interface from multiple aspects, including the task transfer state between computing units, the running state, running record, running monitoring and running log of each computing unit, and the running monitoring of the workflow dimension; workflow result viewing: viewing the running result file in the workflow canvas interface. The technical solution can provide a more convenient, efficient and intuitive workflow management tool for users, reduce computing cost, and support joint computing of computing units of multiple algorithms.

[0005] The above existing technologies all have the following problems: there are limitations in interface interaction; there are limitations in system performance and scalability; and there is a lack of error handling and recovery mechanisms. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention proposes a large-scale model-based intelligent workflow interaction configuration method and system. This method obtains user functional requirements and combines multiple pre-set models, knowledge bases, and plug-ins to construct an overall large-scale model. The large-scale model can output corresponding output requirements based on input requirements and supports knowledge retrieval and configuration. The construction process involves steps such as model selection, knowledge base screening, plug-in integration, and parameter adjustment. The output process includes the generation, evaluation, decision analysis, and optimization of workflow configuration solutions. This invention realizes intelligent and automated workflow configuration, improving configuration efficiency and accuracy.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The intelligent workflow interaction configuration method based on the large model includes:

[0009] Acquire function implementation requirement information input by the first user on the interactive interface, wherein the function implementation requirement information includes input requirement information and corresponding output requirement information;

[0010] According to the function implementation requirement information, combined with multiple preset models, multi-knowledge base retrieval and multi-plug-in integration, build an overall large model to implement the function;

[0011] Based on the overall large model, input demand information of a second user is input, and output demand information of the input demand information is output.

[0012] Specifically, the multiple models include a system reasoning model, an embedding model, and a re-ranking model; the multiple knowledge bases include an industry standard knowledge base, domain knowledge, common sense knowledge, a business rule knowledge base, and a historical data knowledge base; the plug-ins include a data processing plug-in, a format conversion plug-in, and a visualization plug-in.

[0013] Specifically, the specific configuration of the knowledge base search includes:

[0014] A1: Confirm the input requirement information and select the knowledge base to be searched. The knowledge base needs to be created in the system in advance.

[0015] A2: a recall setting is performed, the recall setting is to search all knowledge bases at the same time, retrieve text segments from multiple knowledge bases according to the query intention of the first user, and then select the answer most consistent with the user's question from multiple search results through a reordering process, obtain a sorted and filtered search result list;

[0016] A3: based on the sorted and filtered search result list, according to the business requirements, connect the search results in the search result list to the downstream processing nodes, and configure the processing logic of the downstream nodes, the downstream processing nodes include a question and answer system and a recommendation system.

[0017] Specifically, the construction process of the overall large model includes:

[0018] B1: preprocessing the function implementation requirement information, and according to the preprocessed function implementation requirement information, performing preliminary screening in a plurality of preset basic model libraries, and selecting a basic model;

[0019] B2: analyzing the adaptation degree of the preliminary screened basic model to the current function implementation requirement, generating a basic model candidate list adapted to the current function implementation requirement, and attaching an adaptation analysis report to each basic model in the basic model candidate list;

[0020] B3: analyzing the domain knowledge of the function implementation requirement information, selecting the knowledge base resources highly matched from a plurality of preset knowledge bases according to the determined domain, and performing knowledge extraction and sorting on the screened knowledge base resources, obtaining a domain knowledge graph closely related to the function requirement;

[0021] B4: selecting corresponding plug-ins from a plurality of plug-ins according to the function requirement, and integrating the selected plug-ins, configuring the interface between the plug-ins and the basic model, and generating a plug-in candidate list.

[0022] Specifically, the construction process of the overall large model further includes:

[0023] B5: according to the preprocessed function implementation requirement information and the basic model candidate list, using the domain knowledge graph closely related to the function requirement obtained in B3, combining a back propagation algorithm to adjust the parameters of the basic model, and obtaining a preliminary adjusted basic model;

[0024] B6: using the integrated plug-ins in B4 to adjust and expand the architecture of the preliminary adjusted basic model, reserving an interface for data interaction and function cooperation with the plug-ins, and generating a basic model instance with a plug-in integrated interface;

[0025] B7: According to the base model instance in B6 and the plug-in candidate list in B4, select the plug-in to be integrated, and carry out interface debugging to generate a model system integrated with the plug-in;

[0026] B8: Assemble the base model instance, the domain knowledge graph module and the model system integrated with the plug-in according to business logic and functional flow to form an overall large model realizing the function, and verify the overall large model.

[0027] Specifically, the output process of the overall large model includes:

[0028] C1: Input the preprocessed workflow business requirement description information into the trained large model, and the large model analyzes and reasons the workflow based on the learned knowledge and patterns to generate multiple preliminary workflow configuration schemes;

[0029] C2: According to the pre-set evaluation index system, evaluate the multiple workflow configuration schemes generated by the large model, collect and organize the running results of each workflow configuration scheme in the evaluation tool, and classify and summarize according to the evaluation index system to form an evaluation report of each workflow configuration scheme, the evaluation index system including efficiency index, cost index, quality index, flexibility index and compliance index;

[0030] C3: Use a multi-criteria decision analysis method to assign weights to each evaluation index and calculate the comprehensive score of each configuration scheme;

[0031] C4: Compare the comprehensive scores of all workflow configuration schemes and select the scheme with the highest score as the optimal workflow configuration scheme.

[0032] Specifically, the workflow configuration scheme includes node definition of the workflow, connection relationship between nodes, task allocation rule, condition judgment logic and data transmission path information.

[0033] The intelligent workflow interactive configuration system based on the large model includes a requirement acquisition module, a large model construction module, a knowledge configuration module and a workflow configuration module.

[0034] The requirement acquisition module is configured to acquire and preprocess the function implementation requirement information input by a first user in an interactive interface.

[0035] The large model construction module is configured to construct an overall large model realizing the function according to the preprocessed function implementation requirement information, in combination with a plurality of pre-set models, a plurality of knowledge bases and a plurality of plug-ins.

[0036] The knowledge configuration module is configured to support knowledge retrieval requirements of a second user in a workflow and configure processing logic of downstream nodes according to business requirements.

[0037] The workflow configuration module is configured to input the input requirement information of the second user based on the overall large model and output a workflow configuration scheme.

[0038] Specifically, the large model construction module comprises a model selection unit, a knowledge base screening unit, a plug-in integration unit, a parameter adjustment unit, an assembly and verification unit.

[0039] The model selection unit is configured to preliminarily screen a plurality of preset basic model bases according to the requirement information and select a suitable basic model.

[0040] The knowledge base screening unit is configured to analyze the domain knowledge of the requirement information, screen out a knowledge base resource highly matched therewith from a plurality of preset knowledge bases, perform knowledge extraction and arrangement, and form a domain knowledge graph.

[0041] The plug-in integration unit is configured to select a corresponding plug-in from a plurality of plug-ins according to a functional requirement and integrate and configure an interface between the plug-in and the basic model.

[0042] The parameter adjustment unit is configured to use the domain knowledge graph to adjust the parameters of the basic model in combination with a back propagation algorithm and use the integrated plug-in to adjust and expand the model architecture.

[0043] The assembly and verification unit is configured to finally assemble the basic model instance, the domain knowledge graph module and the model system with the integrated plug-in according to a business logic and a functional flow, form an overall large model, and perform verification.

[0044] Specifically, the knowledge configuration module comprises a recall and sorting unit and a downstream node configuration unit.

[0045] The recall and sorting unit is configured to perform recall setting, search all knowledge bases and re-sort search results according to a query intent to select an answer most conforming to a user question.

[0046] The downstream node configuration unit is configured to connect the search results to a downstream processing node according to a business requirement and configure a processing logic of the downstream node according to a sorted and screened search result list.

[0047] Compared with the prior art, the present application has the following advantages:

[0048] 1. The present application proposes an intelligent workflow interaction configuration system based on a large model and optimizes and improves the architecture, running steps and flow, and the system has the advantages of simple flow, low investment and operation cost and low production cost.

[0049] 2.The application proposes an intelligent workflow interactive configuration method based on large models, which realizes the intelligentization and automation of workflow configuration by constructing an intelligent workflow interactive configuration method based on large models.The method can quickly build an overall large model according to user demand information, combined with multiple models, knowledge bases and plug-ins, and output configuration information that meets user demand based on the model, improving the efficiency of workflow configuration, reducing the tediousness and errors of manual configuration, and improving the accuracy and flexibility of configuration.

[0050] 3.The application proposes an intelligent workflow interactive configuration method based on large models, which further ensures the quality and effect of the workflow configuration scheme through detailed construction and output processes.In the construction process, through the steps of preprocessing of functional demand information, model selection, knowledge base screening and plug-in integration, a domain knowledge graph closely related to functional demand and a plug-in candidate list are formed.In the output process, multiple preliminary workflow configuration schemes are generated, evaluated and weighted, and the optimal configuration scheme is finally selected.This further guarantees the rationality and feasibility of the workflow configuration scheme and improves the performance and efficiency of the entire workflow system. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A schematic diagram of the intelligent workflow interactive configuration method based on large models of the application is shown.

[0052] Figure 2 A schematic diagram of system variable setting of the intelligent workflow interactive configuration method based on large models of the application is shown.

[0053] Figure 3 A schematic diagram of model configuration of the intelligent workflow interactive configuration method based on large models of the application is shown.

[0054] Figure 4 A prompt word arrangement interface diagram of the intelligent workflow interactive configuration method based on large models of the application is shown.

[0055] Figure 5 A system architecture diagram of the intelligent workflow interactive configuration system based on large models of the application is shown. DETAILED DESCRIPTION

[0056] Embodiment 1

[0057] Please refer to Figure 1 The application provides an embodiment: an intelligent workflow interactive configuration method based on large models, including the following steps:

[0058] Obtain the functional implementation demand information input by the first user in the interactive interface, wherein the functional implementation demand information includes input demand information and corresponding output demand information.

[0059] The first user describes the desired function in natural language through a friendly interface, including input data characteristics, format, source, and other input requirement information, as well as the desired output form, content, target, and other output requirement information. For example, the first user may describe, "I need a tool that inputs employee attendance records and outputs monthly attendance statistics for each employee, including the number of days present, the number of tardiness, and the number of early departures."

[0060] Based on the function implementation requirement information, a plurality of models, a plurality of knowledge base retrieval, and a plurality of plug-in integration are combined to construct an overall large model that implements the function.

[0061] Based on the overall large model, input requirement information of a second user is input, and output requirement information of the input requirement information is output.

[0062] In summary, the overall implementation process can be summarized as follows:

[0063] First, after the system is started, the function requirement acquisition stage is entered, and the system provides an intuitive and easy-to-use interface to guide the first user to input function implementation requirement information. The interface can use natural language input box, drop-down menu, example template, and other methods to facilitate the first user to accurately describe their requirements. For example, through the natural language input box, the first user can specify the type of input data, such as text files, database tables, images, and data format requirements, such as CSV format, Excel format, and output result specific forms, such as reports, charts, and text summaries.

[0064] After obtaining the function implementation requirement information of the first user, the system enters the overall model construction stage. The intelligent analysis module built into the system performs semantic analysis and understanding of the requirement information, extracts key requirement characteristics and elements, and then the model selection module selects model components that may be applicable from a plurality of preset model libraries, such as data mining models, prediction models, and classification models. At the same time, the knowledge extraction module searches and extracts knowledge and rules related to the requirements from a plurality of knowledge bases, such as industry standards, business process knowledge, and historical case data. The plug-in management module determines the plug-ins that need to be called according to the requirements, such as data cleaning plug-ins, format conversion plug-ins, and encryption and decryption plug-ins. Finally, the model fusion engine combines and optimizes the selected model components, the knowledge injection mechanism is used to integrate related knowledge into the model, and the plug-in integration framework is used to seamlessly integrate the required plug-ins with the model, and an overall model that can implement the function of the first user is constructed.

[0065] After the overall model is built, the system enters the actual application stage, waiting for the second user to input the demand information. When the second user inputs the relevant data or information according to the input method specified by the overall model, the execution engine of the overall model starts. The execution engine first pre-processes the input information, including data verification, format conversion, data cleaning, etc., to ensure the accuracy and availability of the input information. Then, the pre-processed information is input into the built model, and the model performs calculation and analysis according to the internal algorithm and logic, makes decisions and judgments using the knowledge in the knowledge base, and realizes various specific functions through plug-ins, such as data visualization, file generation, interaction with external systems, etc. Finally, the generated output demand information is presented to the second user in a predetermined way, such as displaying reports on the screen, sending result files by email, storing data into a specified database for subsequent query and use, etc.

[0066] During the implementation of the entire method, the system can also record and analyze the user's operation and feedback to continuously optimize the performance of the model and improve the accuracy and efficiency of the interactive configuration. For example, by analyzing the second user's satisfaction with the output results, usage frequency, modification suggestions, and other feedback information, the system can automatically adjust the parameters of the model, update the knowledge in the knowledge base, and optimize the functions of the plug-ins, so that the overall model can better meet the actual needs of the users, and realize the continuous improvement and optimization of the intelligent workflow interactive configuration.

[0067] The plurality of models include a system reasoning model, an embedding model, and a re-ranking model; the plurality of knowledge bases include an industry standard knowledge base, domain knowledge, common sense knowledge, a business rule knowledge base, and a historical data knowledge base; and the plug-ins include a data processing plug-in, a format conversion plug-in, and a visualization plug-in.

[0068] The specific configuration of the knowledge base retrieval includes:

[0069] A1: Determine the input demand information and select the knowledge base to be retrieved, which needs to be created in the system in advance;

[0070] A2: Perform recall setting, which is to search all knowledge bases at the same time, retrieve text segments from multiple knowledge bases based on the first user's query intent, then select the most suitable answer for the user's question from the multiple search results through the re-ranking process, and obtain a sorted and filtered search result list;

[0071] A3: Based on the sorted and filtered search result list, connect the search results in the search result list to the downstream processing nodes according to the business requirements, and configure the processing logic of the downstream nodes, wherein the downstream processing nodes include a question and answer system and a recommendation system.

[0072] Further, the specific steps of connecting the search results to the downstream processing nodes include:

[0073] (1) Determine the type, format, and quantity of search results, and clarify the type, function, and interface requirements of downstream processing nodes;

[0074] (2) Ensure that the network connection between the search system and the downstream processing nodes is normal, and prepare the necessary data transmission protocols and interface documents;

[0075] (3) According to business needs, sort the search results by relevance, time, score, etc., and filter the search results according to preset filtering conditions such as keyword matching degree, content quality, etc.;

[0076] (4) According to the type and function of the downstream processing node, select the appropriate connection method such as API call, message queue, database sharing, and configure the corresponding interface parameters such as URL, request header, request body according to the interface requirements of the downstream processing node;

[0077] (5) Send the sorted and filtered search results to the downstream processing nodes through the configured interface parameters;

[0078] (6) The downstream processing nodes receive the search results from the search system and perform analysis and processing according to the data format and content of the search results;

[0079] (7) According to business needs, configure the processing logic of the downstream processing nodes such as answer generation of question and answer systems, recommendation algorithm of recommendation systems, etc., and then the downstream processing nodes process the search results according to the configured processing logic and generate corresponding output;

[0080] (8) Monitor the data transmission and processing process between the search system and the downstream processing nodes to ensure the integrity of data transmission and the accuracy of processing, and according to the monitoring results and business needs, optimize and adjust the sorting, filtering, connection method, processing logic, etc. of the search results.

[0081] For example, assuming that the search system has returned a set of search results about "artificial intelligence", the search results need to be connected to the downstream processing nodes of the question and answer system, and the processing logic of the question and answer system needs to be configured: (1) sort the search results by relevance and select the top 10 results most relevant to "artificial intelligence"; (2) send the filtered search results to the interface of the question and answer system through API calls; (3) configure the corresponding processing logic in the question and answer system, such as an answer generation algorithm based on natural language processing technology; (4) After the question and answer system receives the search results, it generates answers related to "artificial intelligence" according to the configured processing logic and returns them to the user. Through the above steps, the search results in the search result list can be effectively connected to the downstream processing nodes, and the corresponding business functions can be realized.

[0082] The construction process of the overall large model includes:

[0083] B1: Preprocess the function implementation requirement information, and based on the preprocessed function implementation requirement information, preliminarily screen in the preset multiple basic model libraries, and select the basic model;

[0084] B2: For the preliminarily screened basic model, analyze its adaptation degree to the current function implementation requirement, generate a basic model candidate list adapted to the current function implementation requirement, and attach an adaptation analysis report to each basic model in each basic model candidate list;

[0085] Further, the specific steps of B2 include:

[0086] (1) Clearly define the functional requirements: understand the specific requirements of the current function implementation, including performance indicators, input and output requirements, data processing capabilities;

[0087] (2) Basic model evaluation: evaluate each of the preliminarily screened basic models one by one, analyze their technical characteristics, performance, and application scope, and during the evaluation process, reference can be made to existing test data, user feedback, and expert opinions;

[0088] (3) Adaptability analysis: analyze the adaptation degree of each basic model to the requirement according to the functional requirement;

[0089] (4) Candidate list generation: according to the adaptability analysis result, list the basic models with higher adaptation degree in the candidate list, and the candidate list can be sorted from high to low according to the adaptation degree for subsequent selection;

[0090] (5) Verification and testing: further verify and test the models in the candidate list to ensure that their performance in actual application meets expectations, and the testing can include unit testing, integration testing, and performance testing;

[0091] (6) Final selection: based on the test results, select the base model that best fits the current functional implementation requirements.

[0092] B3: Analyze the domain knowledge of the functional implementation requirement information, select the knowledge base resources that match the determined domain from multiple preset knowledge bases, and perform knowledge extraction and organization on the selected knowledge base resources to obtain a domain knowledge graph closely related to the functional requirements;

[0093] Further, the specific steps of B3 include:

[0094] (1) Functional requirement analysis and domain determination: Clearly define the specific domain knowledge required for functional implementation, and determine the core concepts and scope of the domain knowledge;

[0095] (2) Knowledge base resource selection: From multiple preset knowledge bases, select according to the core concepts and scope of the domain knowledge, evaluate the matching degree of the knowledge base resources and the functional requirements, and select the knowledge base resources with high matching degree;

[0096] (3) Knowledge extraction: Use rule-based knowledge extraction methods to extract knowledge from the selected knowledge base resources, including entity extraction, relationship extraction, and attribute extraction. The rule-based knowledge extraction method is a prior art in the field and is not part of the inventive concept of this application, so it will not be described here;

[0097] (4) Knowledge organization and graph construction: Organize the extracted knowledge to form a structured knowledge representation, and construct a domain knowledge graph based on the knowledge representation, including the definition and connection of nodes, relationships, attributes, and other elements;

[0098] (5) Knowledge graph verification and optimization: Verify the constructed knowledge graph to ensure the accuracy and completeness of the knowledge, and optimize and adjust the knowledge graph based on the verification results.

[0099] B4: According to the functional requirements, select the corresponding plug-in from multiple plug-ins, and integrate the selected plug-in, configure the interface between the plug-in and the base model, and generate a plug-in candidate list;

[0100] B5: Based on the preprocessed functional implementation requirement information and the base model candidate list, use the domain knowledge graph closely related to the functional requirements obtained in B3, and combine the backpropagation algorithm to adjust the parameters of the base model to obtain a preliminary adjusted base model;

[0101] Further, the specific steps of B5 include:

[0102] (1) Data preprocessing: Clean, standardize, and normalize the function implementation requirement information to ensure data quality and consistency, and convert the processed data into a format acceptable to the model;

[0103] (2) Constructing a domain knowledge graph: Using the domain knowledge closely related to functional requirements obtained in B3, construct or improve the knowledge graph to ensure that the entities, relationships, and attributes in the knowledge graph are relevant to functional requirements;

[0104] (3) Select the basic model: From the list of basic model candidates, select the most appropriate model as the starting point based on the functional requirements and the characteristics of the domain knowledge graph;

[0105] (4) Initializing model parameters: Randomly initializing the parameters of the selected basic model. Randomly initializing parameters is a prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0106] (5) Combined with knowledge graph for forward propagation: input the preprocessed data into the model, and use the information in the domain knowledge graph to assist the model in forward propagation calculation;

[0107] (6) Calculating the loss function: Calculating the loss function based on the model output and the true label. The loss function calculation formula is the existing technical content in this field and is not an inventive solution of this application, so it is not described here;

[0108] (7) Back propagation and parameter update: Using the back propagation algorithm, the gradient is calculated layer by layer from the output layer to the input layer, and the SGD optimization algorithm is used to update the model parameters based on the gradient calculation results. The back propagation algorithm and the SGD optimization algorithm are prior art contents in this field and are not the inventive solution of this application, and are not described in detail here;

[0109] (8) Iterative training: Repeat steps (5) to (7) until the model performance on the validation set reaches the preset stopping condition;

[0110] (9) Obtaining the preliminarily adjusted basic model: After multiple iterative trainings, the preliminarily adjusted basic model is obtained.

[0111] B6: Use the plug-ins integrated in B4 to adjust and expand the architecture of the initially adjusted basic model, reserve interfaces for data interaction and functional collaboration with the plug-ins, and generate a basic model instance with a plug-in integration interface.

[0112] B7: Based on the base model instance in B6 and the plugin candidate list in B4, select the plugin to be integrated, and perform interface debugging to generate a model system with integrated plugins. Interface debugging is a prior art in the field and is not part of the inventive concept of this application, so it is not described here.

[0113] B8: Assemble the base model instance, the domain knowledge graph module, and the model system with integrated plugins according to business logic and functional flow to form a whole large model that implements the function, and verify the whole large model.

[0114] Further, the specific steps of B8 include:

[0115] (1) Component preparation: Ensure that the base model instance has been trained and meets the expected performance, the domain knowledge graph module has been built and contains entities, relationships and attributes closely related to the functional requirements, and the model system with integrated plugins has been developed and has passed preliminary testing;

[0116] (2) According to the business logic and functional flow, design the architecture of the whole large model, determine the interface and data interaction method between components;

[0117] (3) Component integration: Integrate the base model instance, the domain knowledge graph module, and the model system with integrated plugins according to the designed architecture to ensure smooth data interaction and interface matching between components;

[0118] (4) Function test: Test the function of the whole large model to verify whether it meets the functional requirements;

[0119] (5) Record the problems found during testing and fix them, and optimize the performance of the whole large model according to the test results.

[0120] Embodiment 2

[0121] Please refer to Figures 2-4 The output process of the whole large model in this embodiment includes:

[0122] C1: Input the preprocessed workflow business requirement description information into the trained large model. The large model analyzes and reasons the workflow based on the knowledge and patterns it has learned, generating multiple preliminary workflow configuration schemes;

[0123] Wherein, the workflow simplifies the complexity of the system by dividing complex tasks into a series of smaller, easier-to-manage steps, reducing the dependence on prompt words and model reasoning ability, enhancing the performance in handling complex tasks, and improving the system's explainability, stability and fault tolerance. Specifically, it includes:

[0124] (1) Workflow creation and start: To start a workflow, you can choose to create a new workflow from scratch, which requires some basic operation skills, such as how to add nodes in the workspace, how to connect nodes to each other, and how to configure nodes, as well as how to debug and view the running record of the workflow; after completing the construction of the workflow, save and publish the workflow so that it can be used; finally, execute the workflow through the published application.

[0125] (2) Variables: Workflow applications are composed of multiple independent nodes, most of which have input and output items, but the required input information and output results of each node are different. In order to represent dynamic changes with a fixed symbol, variables are used as dynamic data containers to store and transfer variable information and reference each other between different nodes, thereby achieving flexible communication.

[0126] (3) Nodes, nodes are the basic elements of building a workflow. By connecting nodes with different functions, you can complete the continuous operation in the workflow, including:

[0127] 1) In each workflow application, there is a preset node called "Start", which provides the necessary initial information and data for the start of the entire workflow, such as user input information, to ensure that the workflow can proceed smoothly. At the same time, in the settings page of the start node, you can see two parts of the settings, which are input fields and preset system variables, as shown in Figure 2

[0128] 2) Use LLM nodes to process user input information in nodes. Use its dialogue, generation, classification, and processing capabilities in the workflow to efficiently handle various tasks, and can be applied in different parts of the workflow. The configuration steps of LLM include:

[0129] a. In the application editing page, you can add LLM nodes by clicking the right mouse button, dragging the + sign at the end of the previous node, or dragging the LLM on the left side;

[0130] b Configure model parameters, which will affect the output results of the model, such as temperature, TopP, maximum token number, and reply format. The system provides three sets of preset parameters in this invention: creative, balanced, and accurate. If you are not familiar with these parameters, you can choose the default settings, as shown in Figure 3

[0131] c. Write prompt words, including system and user parts, as shown in Figure 4 ​​As shown, in the prompt word editor, you can enter " / " or click "(x)" to call out the variable insertion menu and insert the upstream node variable into the prompt word as context content;

[0132] 3) Retrieve text content related to the user's question from the knowledge base.

[0133] C2: Evaluate multiple workflow configuration solutions generated by the large model based on a pre-defined evaluation index system. At the same time, collect and organize the running results of each workflow configuration solution in the evaluation tool, and classify and summarize them according to the evaluation index system to form an evaluation report for each workflow configuration solution. The evaluation index system includes but is not limited to efficiency indicators, cost indicators, quality indicators, flexibility indicators, and compliance indicators.

[0134] C3: Use multi-criteria decision analysis methods Assign weights to each evaluation indicator and calculate the comprehensive score of each configuration scheme , where A represents the judgment matrix constructed after comparing and scoring the evaluation indicators pairwise. represents the maximum eigenvalue of the judgment matrix, Indicates the j The comprehensive score of the workflow configuration scheme, w represents the weight, m represents the number of evaluation indicators, Indicates the i The weights assigned to the evaluation indicators are: Indicates the j The workflow configuration scheme is in i Quantitative scores on the evaluation indicators;

[0135] It should be noted that A is a The matrix, is a scalar, w is a vector.

[0136] C4: Compare the comprehensive scores of all workflow configuration schemes and select the scheme with the highest score as the optimal workflow configuration scheme.

[0137] The workflow configuration scheme includes workflow node definitions, connection relationships between nodes, task allocation rules, conditional judgment logic, and data transmission path information.

[0138] Example 2

[0139] See also Figure 5 Another embodiment provided by the present invention is an intelligent workflow interactive configuration system based on a large model, comprising:

[0140] The demand acquisition module, the large model construction module, the knowledge configuration module, and the workflow configuration module;

[0141] The demand acquisition module is configured to acquire and preprocess the function implementation demand information input by the first user on the interactive interface, thereby providing a basis for subsequent large model construction.

[0142] The large model construction module is configured to construct an overall large model for implementing the function according to the preprocessed function implementation demand information, in combination with a plurality of preset models, a plurality of knowledge bases, and a plurality of plug-ins.

[0143] The knowledge configuration module is configured to support the knowledge retrieval demand of the second user in the workflow and configure the processing logic of the downstream nodes according to the business demand.

[0144] The workflow configuration module is configured to input the input demand information of the second user based on the overall large model and output a workflow configuration scheme, while supporting the evaluation and optimization of the configuration scheme.

[0145] The large model construction module includes a model selection unit, a knowledge base screening unit, a plug-in integration unit, a parameter adjustment unit, and an assembly and verification unit.

[0146] The model selection unit is configured to preliminarily screen the demand information in a plurality of preset basic model bases and select an appropriate basic model.

[0147] The knowledge base screening unit is configured to analyze the domain knowledge of the demand information, screen out high-matching knowledge base resources from a plurality of preset knowledge bases, perform knowledge extraction and organization, and form a domain knowledge graph.

[0148] The plug-in integration unit is configured to select corresponding plug-ins from a plurality of plug-ins according to the function demand and integrate and configure the interface between the plug-ins and the basic model.

[0149] The parameter adjustment unit is configured to use the domain knowledge graph to adjust the parameters of the basic model in combination with a back propagation algorithm, and use the integrated plug-ins to adjust and expand the model architecture.

[0150] The assembly and verification unit is configured to finally assemble the basic model instance, the domain knowledge graph module, and the model system with integrated plug-ins according to the business logic and function flow, form an overall large model, and perform verification.

[0151] The knowledge configuration module includes a recall and sorting unit and a downstream node configuration unit.

[0152] The recall and sorting unit is configured to perform recall setting, search all knowledge bases, reorder the search results according to the query intent, and select the answers most consistent with the user's questions.

[0153] The downstream node configuration unit is configured to connect the search results to the downstream processing nodes according to the service requirements based on the sorted and filtered search result list, and configure the processing logic of the downstream nodes.

[0154] The workflow configuration module comprises a workflow generation unit, an evaluation unit, a decision analysis unit and an optimization selection unit.

[0155] The workflow generation unit is configured to input the preprocessed workflow service requirement description information into the trained large model to generate a plurality of preliminary workflow configuration schemes.

[0156] The evaluation unit is configured to evaluate the plurality of workflow configuration schemes generated by the large model according to a pre-set evaluation index system to form an evaluation report.

[0157] The decision analysis unit is configured to perform weight distribution on each evaluation index by using a multi-criteria decision analysis method to calculate a comprehensive score of each configuration scheme.

[0158] The optimization selection unit is configured to compare the comprehensive scores of all workflow configuration schemes and select the scheme with the highest score as the optimal workflow configuration scheme.

[0159] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative and not restrictive, and a person of ordinary skill in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments without departing from the purpose of the present application and the scope of protection, and these are all within the protection of the present application.

[0160] If the technical solution of the present disclosure involves personal information, the product applying the technical solution of the present disclosure has explicitly informed the personal information processing rules before processing the personal information and obtained the personal independent consent. If the technical solution of the present disclosure involves sensitive personal information, the product applying the technical solution of the present disclosure has obtained the personal independent consent before processing the sensitive personal information, and at the same time meets the requirement of "explicit consent". For example, at the personal information collection device such as camera, a clear and prominent sign is set to inform that the personal information collection range has been entered and the personal information will be collected, and if the individual voluntarily enters the collection range, it is considered to agree to collect the personal information; or on the device for processing personal information, the personal information processing rules are informed by using obvious signs / information, and the personal authorization is obtained by pop-up information or asking the individual to upload the personal information; wherein, the personal information processing rules can include personal information processor, personal information processing purpose, processing method and personal information type, etc.

Claims

1. An intelligent workflow interaction configuration method based on a large model, characterized in that: include: Acquire function implementation requirement information input by the first user on the interactive interface, wherein the function implementation requirement information includes input requirement information and corresponding output requirement information; According to the function implementation requirement information, combined with multiple preset models, multi-knowledge base retrieval and multi-plug-in integration, build an overall large model to implement the function; Based on the overall large model, inputting input demand information of a second user, and outputting output demand information of the input demand information; The construction process of the overall large model includes: B1: Preprocess the function realization requirement information, and conduct preliminary screening in multiple preset basic model libraries based on the preprocessed function realization requirement information to select the basic model; B2: Analyze the adaptability of the preliminarily selected basic models to the current functional implementation requirements, generate a candidate list of basic models that are suitable for the current functional implementation requirements, and provide an adaptability analysis report for each basic model in the candidate list; B3: Analyze the domain knowledge of the functional implementation requirements information. Based on the determined domain, select highly matching knowledge base resources from multiple preset knowledge bases. Extract and organize the selected knowledge base resources to obtain a domain knowledge graph closely related to the functional requirements. B4: Select corresponding plug-ins from multiple plug-ins based on functional requirements, integrate the selected plug-ins, configure the interface between the plug-ins and the basic model, and generate a plug-in candidate list; The construction process of the overall large model also includes: B5: Based on the pre-processed functional implementation requirement information and the basic model candidate list, use the domain knowledge graph closely related to the functional requirements obtained in B3 and the back-propagation algorithm to adjust the parameters of the basic model to obtain a preliminarily adjusted basic model. B6: Use the plug-ins integrated in B4 to adjust and expand the architecture of the initially adjusted basic model, reserve interfaces for data interaction and functional collaboration with the plug-ins, and generate a basic model instance with a plug-in integration interface. B7: Based on the basic model instance in B6 and the plug-in candidate list in B4, select the plug-ins to be integrated and perform interface joint debugging to generate a model system with integrated plug-ins. B8: The basic model instance, domain knowledge graph module and the model system with integrated plug-ins are finally assembled according to the business logic and functional process to form an overall large model that realizes the functions, and the overall large model is verified.

2. The method for configuring intelligent workflow interactions based on a large model according to claim 1, wherein: The multiple models include a system reasoning model, an embedding model, and a re-ranking model; the multiple knowledge bases include an industry standard knowledge base, domain knowledge, common sense knowledge, a business rule knowledge base, and a historical data knowledge base; the plug-ins include a data processing plug-in, a format conversion plug-in, and a visualization plug-in.

3. The method for configuring intelligent workflow interaction based on a large model according to claim 2, wherein: The specific configuration of the knowledge base retrieval includes: A1: Confirm the input requirement information and select the knowledge base to be searched. The knowledge base needs to be created in the system in advance. A2: Performing a recall setting, which searches all knowledge bases simultaneously. Based on the first user's query intent, text snippets are retrieved from multiple knowledge bases. Then, through a re-ranking process, the answer that best matches the user's question is selected from the multiple search results to obtain a sorted and filtered search result list. A3: Based on the sorted and filtered search results list, according to business needs, the search results in the search results list are connected to downstream processing nodes. At the same time, the processing logic of the downstream nodes is configured. The downstream processing nodes include the question-answering system and the recommendation system.

4. The method for configuring intelligent workflow interaction based on a large model according to claim 3, wherein: The output process of the overall large model includes: C1: Input the pre-processed workflow business requirement description information into the trained large model. Based on its learned knowledge and patterns, the large model analyzes and infers the workflow and generates multiple preliminary workflow configuration solutions. C2: Evaluate multiple workflow configuration solutions generated by the large model based on a pre-defined evaluation index system. At the same time, collect and organize the running results of each workflow configuration solution in the evaluation tool, classify and summarize them according to the evaluation index system to form an evaluation report for each workflow configuration solution. The evaluation index system includes efficiency indicators, cost indicators, quality indicators, flexibility indicators, and compliance indicators. C3: Use multi-criteria decision analysis to assign weights to each evaluation indicator and calculate the comprehensive score of each configuration scheme; C4: Compare the comprehensive scores of all workflow configuration schemes and select the scheme with the highest score as the optimal workflow configuration scheme.

5. The method for configuring intelligent workflow interaction based on a large model according to claim 4, characterized in that: The workflow configuration scheme includes workflow node definitions, connection relationships between nodes, task allocation rules, conditional judgment logic, and data transmission path information.

6. A large-model-based intelligent workflow interactive configuration system, which is used to implement the large-model-based intelligent workflow interactive configuration method according to any one of claims 1 to 5, characterized in that: include: Requirements acquisition module, large model building module, knowledge configuration module, workflow configuration module; The requirement acquisition module is used to acquire and pre-process the function implementation requirement information input by the first user in the interactive interface; The large model construction module is used to construct an overall large model that realizes the function based on the pre-processed function realization requirement information, combined with multiple preset models, multiple knowledge bases and multiple plug-ins; The knowledge configuration module is used to support the knowledge retrieval needs of the second user in the workflow and configure the processing logic of the downstream nodes according to business needs; The workflow configuration module is used to input the input requirement information of the second user based on the overall large model and output a workflow configuration solution.

7. The intelligent workflow interactive configuration system based on a large model according to claim 6, characterized in that: The large model construction module includes: a model selection unit, a knowledge base screening unit, a plug-in integration unit, a parameter adjustment unit, and an assembly and verification unit; The model selection unit is used to perform preliminary screening in a plurality of preset basic model libraries according to the demand information and select an adapted basic model; The knowledge base screening unit is used to analyze the domain knowledge of the demand information, screen out knowledge base resources that are highly matched with it from multiple preset knowledge bases, and extract and organize the knowledge to form a domain knowledge graph; The plug-in integration unit is used to select a corresponding plug-in from multiple plug-ins according to functional requirements, integrate the plug-ins, and configure the interface between the plug-in and the basic model; The parameter adjustment unit is used to use the domain knowledge graph in combination with the back propagation algorithm to adjust the parameters of the basic model, and use the integrated plug-in to adjust and expand the model architecture; The assembly and verification unit is used to finally assemble the basic model instance, domain knowledge graph module and the model system with integrated plug-ins according to the business logic and functional process to form an overall large model and perform verification.

8. The intelligent workflow interactive configuration system based on a large model according to claim 7, characterized in that: The knowledge configuration module includes: a recall and sorting unit and a downstream node configuration unit; The recall and sorting unit is used to perform recall settings, search all knowledge bases, and re-sort the search results according to the query intent to select the answer that best meets the user's question; The downstream node configuration unit is used to connect the search results to the downstream processing nodes according to the sorted and filtered search result list and business needs, and configure the processing logic of the downstream nodes.

Images