Application method for realizing a travel enterprise assistant based on a large model

Abstract

The present invention discloses an application method for realizing a travel enterprise assistant based on a large model, which includes obtaining the input data and user depth information of a user through a front-end interface, and saving the user depth information into a custom variable; retrieving a static database based on workflow orchestration technology according to the input data to obtain a static output result; establishing a logical address mapping relationship between the input data and the content of a dynamic database according to the large language inference model, outputting a dynamic output result, combining the dynamic output result with the static data result to form an application output result, establishing a problem classifier, selecting an output method according to the classification situation of the problem classifier for the application output result, and displaying it on the front-end interface. The present invention is based on a reinforcement learning mechanism of user stay duration and secondary follow-up rate, enabling the model to adaptively optimize, continuously improving user satisfaction, and the problem classifier dynamically selects the output form according to user preferences, with the information transmission efficiency increased by 40%.

Description

Technical Field

[0001] The present invention relates to the technical field of data search, and in particular to an application method for realizing a travel enterprise assistant based on a large model. Background Art

[0002] At present, there are some problems in the technical field of data search:

[0003] 1. Processing capacity limitation: Traditional enterprise systems often rely on rule-driven algorithms, which have limited capabilities in processing complex tasks such as natural language understanding, image recognition, or predictive analysis. They cannot learn patterns and features from large amounts of data like large models.

[0004] 2. Lack of context understanding: Traditional systems usually cannot understand complex context information, resulting in an inability to provide accurate and in-depth answers or suggestions when processing user queries or making automated decisions. Summary of the Invention

[0005] In view of the problems existing in the existing application method for realizing a travel enterprise assistant based on a large model, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is: to provide a new enterprise assistant application module, which integrates a large pre-trained model to enhance the processing ability for complex tasks.

[0007] To solve the above technical problems, the present invention provides the following technical solution: an application method for realizing a travel enterprise assistant based on a large model, which includes,

[0008] Obtain the input data and user depth information of the user through the front-end interface, save the user depth information into a custom variable, where the input data includes natural language, uploaded files, and pictures, and the user depth information includes the semantic complexity of the user input, the residence time of the user interface, and the page secondary questioning rate;

[0009] Retrieve the static database based on the input data according to the workflow orchestration technology to obtain a static output result;

[0010] Configure the model parameters according to the custom variable to establish a large language inference model;

[0011] Establish a logical address mapping relationship between the input data and the content of the dynamic database according to the large language inference model, and output a dynamic output result;

[0012] Combine the dynamic output result with the static data result to form an application output result;

[0013] A problem classifier is established to select an output method for the output result of the application according to the classification of the problem classifier. The problem classifier refers to the user's selection of the output method for the application output result, and the output method selection includes text output, generating charts, and exporting reports.

[0014] Output the output result of the application through the classification method output by the problem classifier and display it on the front-end interface.

[0015] As a preferred solution of the application method for implementing a tourism enterprise assistant based on a large model according to the present invention, wherein, obtaining the user's in-depth information through the front-end interface includes:

[0016] Obtain the semantic complexity, user stay duration, and secondary questioning rate when the user inputs at the front end.

[0017] The obtaining of the semantic complexity when the user inputs at the front end includes:

[0018] Judge the depth of dependency syntactic analysis and the density of correlation words in the semantic complexity. When the depth of dependency syntactic analysis exceeds the threshold or the keyword density is lower than the threshold, the system starts a three-level retrieval mechanism: the first layer uses semantic embedding matching based on BERT, the middle layer uses an improved TF-IDF weighted algorithm, and the bottom layer triggers cross-domain knowledge graph association. After the user's retrieval, dynamically adjust the model parameters through the user stay duration and secondary questioning rate to form a reinforcement learning mechanism with a time decay factor.

[0019] The dynamically adjusting the model parameters through the user stay duration and secondary questioning rate includes that when the user stay duration exceeds the stay duration threshold and the secondary questioning rate exceeds the secondary questioning rate threshold, a reinforcement learning mechanism with a time decay factor is formed to dynamically reduce the model parameters in the large language inference model.

[0020] As a preferred solution of the application method for implementing a tourism enterprise assistant based on a large model according to the present invention, wherein, when the depth of dependency syntactic analysis exceeds the threshold or the keyword density is lower than the threshold, the system starting the three-level retrieval mechanism includes:

[0021] The first layer uses semantic embedding matching based on BERT, the middle layer uses an improved TF-IDF weighted algorithm, introducing a word vector cosine similarity correction factor α = 0.65, and the bottom layer triggers cross-domain knowledge graph association.

[0022] The depth of dependency syntactic analysis exceeding the threshold or the keyword density being lower than the threshold includes:

[0023] When the depth of dependency syntactic analysis ≥ 5 layers or the keyword density < 30%, trigger the retrieval mechanism.

[0024] As a preferred solution of the application method for implementing a travel enterprise assistant based on a large model according to the present invention, wherein, the retrieving the static database based on the workflow orchestration technology according to the input data to obtain a static output result includes:

[0025] When the application module receives a data stream exceeding the input data threshold, the fast sorting channel is automatically activated, and the data stream refers to the number of user access requests;

[0026] Quickly extract key features according to the hardware acceleration technology, and the key features include data fluctuation patterns and time sensitivity;

[0027] Perform rough classification according to the key feature multi-layer filtering screen. The first layer of the rough classification screens out obviously irrelevant data, the second layer pre-groups according to domain labels, and the third layer performs priority sorting based on historical pattern matching.

[0028] As a preferred solution of the application method for implementing a travel enterprise assistant based on a large model according to the present invention, wherein, outputting the content within the logical address range in the database to the front-end interface, the content output includes reply text, generating charts, and exporting reports, and the content output method is freely selected by the user during the content input process;

[0029] The database includes a static database and a dynamic database called and generated instantaneously;

[0030] The content output method includes selecting the application output result formed by combining the dynamic output result and the static data result as the content output.

[0031] As a preferred solution of the application method for implementing a travel enterprise assistant based on a large model according to the present invention, wherein, the process of forming the reply text includes:

[0032] The user inputs data, and the application interface obtains the user input through the sys.query variable. After obtaining the application output result, use the question classifier to judge whether the content output method is reply text, generating charts, or exporting reports; if it is judged to display text, directly reply the text to the front-end interface; if it is judged to generate charts or export reports, stop the output and return to the question classification stage.

[0033] As a preferred solution of the application method for implementing a travel enterprise assistant based on a large model according to the present invention, wherein, the process of generating the chart includes:

[0034] The user inputs data. The application interface obtains the user input through the sys.query variable. After obtaining the application output result, a question classifier is used to determine whether the content output method is to reply with text, generate a chart, or export a report. If it is determined to generate a chart, the parameter extractor is used to obtain the x-axis data type and y-axis data type of the chart from sys.query, and the application output result is input into the x-axis and y-axis to generate the x and y-axis data of the chart. Then, the bar chart tool node is called to generate the chart to the front-end interface. If it is determined to reply with text or export a report, the output is stopped and the process returns to the question classification stage.

[0035] As a preferred solution of the application method for implementing a travel-related enterprise assistant based on a large model according to the present invention, wherein the process of forming the exported report includes:

[0036] The user inputs data. The application interface obtains the user input through the sys.query variable. After obtaining the application output result, a question classifier is used to determine whether the content output method is to reply with text, generate a chart, or export a report. If it is determined to export a report, the parameter extractor is used to obtain the horizontal data type and vertical data type of the report from sys.query, and the report tool is called to generate a report. The application output result is input into the horizontal data type and vertical data type to generate the report and display it on the front-end interface. If it is determined to reply with text or generate a chart, the output is stopped and the process returns to the question classification stage.

[0037] The present invention provides the following technical solution: An electronic device, comprising:

[0038] One or more processors;

[0039] A storage device having one or more programs stored thereon;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement a method for identifying data violation operation behaviors in an AI-based business system.

[0041] The present invention provides the following technical solution: An electronic device, comprising:

[0042] A computer-readable storage medium having executable instructions stored thereon, and when the instructions are executed by a processor, the processor implements a method for identifying data violation operation behaviors in an AI-based business system.

[0043] Advantages of the present invention

[0044] 1. Enhanced processing capabilities: Provide a new type of enterprise assistant application module that integrates large pre-trained models to enhance the processing capabilities for complex tasks, especially in natural language understanding, image recognition, and predictive analysis, enabling it to learn deep patterns and features from large amounts of data.

[0045] 2. Improved context understanding capabilities: Develop an enterprise assistant application module that can understand and process complex context information to ensure more accurate and in-depth answers or suggestions are provided when handling user queries or making automated decisions.

[0046] 3. Optimized data utilization: Design a mechanism to fully exploit the potential value of a large amount of underutilized data within the enterprise using large pre-trained models to enhance the effectiveness and efficiency of data-driven decision-making.

[0047] 4. Increased response speed and accuracy: By integrating large models, achieve fast and accurate responses to complex or ambiguous user queries, improving the user experience and service quality, especially in the fields of customer service and data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the implementation method for the application method of the travel-related enterprise assistant based on the large model in Embodiment 1;

[0049] Figure 2 It is an architecture diagram of the enterprise assistant for the application method of the travel-related enterprise assistant based on the large model in Embodiment 2;

[0050] Figure 3 It is a schematic diagram of the connections between modules for the application method of the travel-related enterprise assistant based on the large model in Embodiment 2;

[0051] Figure 4 It is a schematic diagram of the electronic structure for the application method of the travel-related enterprise assistant based on the large model in Embodiment 4. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made in conjunction with the accompanying drawings of the specification.

[0053] Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0054] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other with other embodiments.

[0055] Embodiment 1

[0056] Referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides an application method for realizing a travel enterprise assistant based on a large model, which includes:

[0057] Obtain the input data and user depth information of the user through the front-end interface, save the user depth information into a custom variable. The input data includes natural language, uploaded files, and pictures, and the user depth information includes the semantic complexity of the user input, the dwell time on the user interface, and the page secondary inquiry rate;

[0058] Retrieve the static database based on the input data using workflow orchestration technology to obtain a static output result;

[0059] Configure model parameters according to the custom variable to establish a large language inference model;

[0060] Establish a logical address mapping relationship between the input data and the content of the dynamic database according to the large language inference model, and output a dynamic output result;

[0061] Combine the dynamic output result and the static data result to form an application output result;

[0062] Establish a question classifier, and select an output method for the application output result according to the classification of the question classifier. The question classifier refers to the user's selection of the output method for the application output result, and the output method selection includes text output, generating charts, and exporting reports;

[0063] Output the application output result in the classification method output by the question classifier and display it in the front-end interface.

[0064] Judge the dependency syntactic analysis depth and conjunction density in the semantic complexity. When the dependency syntactic analysis depth ≥ 5 layers or the keyword density < 30%, the system starts a three-level retrieval mechanism: the first layer uses semantic embedding matching based on BERT, the middle layer uses an improved TF-IDF weighted algorithm, introduces a word vector cosine similarity correction factor α = 0.65, and the bottom layer triggers cross-domain knowledge graph association. When the user dwell time exceeds the dwell time threshold of 8s and the secondary inquiry rate exceeds the secondary inquiry rate threshold of 40%, a reinforcement learning mechanism with a time decay factor is formed to dynamically reduce the model parameters in the large language inference model.

[0065] Dynamically adjusting model parameters based on user dwell time and secondary inquiry rate includes forming a reinforcement learning mechanism with a time decay factor and dynamically reducing the model parameters in the large language inference model when the user dwell time exceeds the dwell time threshold and the secondary inquiry rate exceeds the secondary inquiry rate threshold.

[0066] Retrieving static database based on input data using workflow orchestration technology to obtain static output results includes automatically starting a fast sorting channel when the application module receives a data stream exceeding the input data threshold, where the data stream refers to the number of user access requests.

[0067] Quickly extracting key features according to hardware acceleration technology, where the key features include data fluctuation patterns and time sensitivity.

[0068] Performing rough classification according to the key feature multi-layer filtering screen, where the first layer of the rough classification screens out obviously irrelevant data, the second layer pre-groups by domain labels, and the third layer performs priority sorting based on historical pattern matching.

[0069] Outputting the content within the logical address range in the database to the front-end interface, where the content output includes reply text, generating charts, and exporting reports, and the content output method is freely selected by the user during the content input process.

[0070] The database includes a static database and a dynamic database that is called and generated immediately.

[0071] The process of forming the reply text is as follows:

[0072] The user inputs data, obtains the user input through the sys.query variable, inputs sys.query as a parameter into the static database for retrieval, inputs the retrieval result into the large language inference model, calls the large language inference model to retrieve the output result of the dynamic database according to the large language inference model prompt words and the large language inference model type, then re-organizes the static database retrieval content and the dynamic database retrieval content according to the content input by the user, and uses a question classifier to judge whether it is reply text, generating charts, or exporting reports; if it is judged as displaying text, directly reply the text to the front-end interface; if it is judged as generating charts or exporting reports, stop the output and return to the question classification stage.

[0073] The process of forming the generated chart is as follows:

[0074] The user inputs data, obtains the user input through the sys.query variable, uses sys.query as a parameter to retrieve in the static database, inputs the retrieval result into the large language inference model, and according to the large language inference model prompt words and the large language inference model type, calls the large language inference model to retrieve the dynamic database to output the result. Then, according to the content input by the user, combines the static database retrieval content and the dynamic database retrieval content to generate an application output result. Use a question classifier to determine whether to reply with text, generate a chart, or export a report; if it is determined to generate a chart, obtain the x-axis data type and y-axis data type of the chart from sys.query through a parameter extractor, input the application output result into the x-axis and y-axis, generate the x and y-axis data of the chart, and call the bar chart tool node to generate the chart on the front-end interface; if it is determined to reply with text or export a report, stop the output and return to the question classification stage.

[0075] The formation process of exporting a report is as follows:

[0076] The user inputs data, obtains the user input through the sys.query variable, uses sys.query as a parameter to retrieve in the static database, inputs the retrieval result into the large language inference model, and according to the large language inference model prompt words and the large language inference model type, calls the large language inference model to retrieve the dynamic database to output the result. Then, according to the content input by the user, combines the static database retrieval content and the dynamic database retrieval content to generate an application output result. Use a question classifier to determine whether to reply with text, generate a chart, or export a report; if it is determined to export a report, obtain the horizontal data type and vertical data type of the report from sys.query through a parameter extractor, call the report tool to generate a report, input the application output result into the horizontal data type and vertical data type, and generate the report to be displayed on the front-end interface; if it is determined to reply with text or generate a chart, stop the output and return to the question classification stage.

[0077] Output the application output result through the classification method output by the question classifier and display it on the front-end interface.

[0078] Embodiment 2

[0079] In the second embodiment of the present invention, which is different from the first embodiment: in the previous embodiment, the application method of realizing the tourism-related enterprise assistant based on the large model includes:

[0080] On the basis of building smart tourism, tourism-related enterprises provide a chat-based window through the application method of realizing the tourism-related enterprise assistant based on the large model, allowing enterprise users to obtain the data they want through this front-end function, including basic data precipitated over time, dynamic data generated immediately, statistical data obtained through analysis, etc.

[0081] The applications include functions such as enterprise assistants, intelligent customer service, and financial statements, and can be further expanded according to requirements in the future.

[0082] Establish an application architecture: Based on large models, implement the application method architecture of travel-related enterprise assistants, which is divided into the application front-end, intelligent agent, knowledge base, database, and large model.

[0083] Application front-end: A program that provides an operable interface for enterprise employees to operate, interacts with the intelligent agent of the travel-related enterprise assistant through the front-end interface, and obtains the desired data;

[0084] Intelligent agent: Through the method of workflow orchestration, combines the emergent capabilities of large models by using static knowledge base data and calling dynamic database data to implement the functions of enterprise assistants;

[0085] Knowledge base / database: Provide comprehensive enterprise data to users by importing static data into the knowledge base and calling instantaneously generated dynamic data;

[0086] Large model: Call the general large model provided by the platform side and the open-source large model deployed locally.

[0087] Application architecture Figure 2 Is as follows.

[0088] When the relationship between modules starts to be called, the user accesses the application front-end, enters content in the chat window of the front-end, and the intelligent agent calls the knowledge base and database to achieve real-time access to local data, and then submits the obtained content to the large model for processing or prediction, and finally outputs the result to the user. The call relationship between modules is as Figure 3 Shown.

[0089] Processing flow: The user enters the required question in the front-end interface. Through the "question classification" node, the large model classifies the question. The questions are divided into three categories: reply text, generate charts, and export reports. The categories can be expanded according to business requirements, and different categories have dedicated processing flows.

[0090] Reply text: Enter the reply text process. First, use the "knowledge retrieval" node of the local knowledge base to recall relevant content in the knowledge base, then enter the "call large model" node. The large model infers and outputs the result based on the knowledge base recalled by the user's question, and then through the "question classification" node, it is judged whether to display as text or in a list, and the "reply" node is called to display the list or a single text on the front-end respectively.

[0091] Generate Chart: Enter the chart generation process. First, classify the question through the "Question Classification" node. The questions are classified into "Query Passenger Flow Data", "Query Ticket Data", "Query Hotel Data", and "Query Scenic Spot Data". Then, simultaneously call the "Knowledge Retrieval" node to recall relevant knowledge and call the interface through the "Http Request" node to obtain data from the database. Input the data from both into the "Large Model" node. The large model performs reasoning and gives the response content. Finally, call the "Chart Generation Tool" to organize the data into a chart, and return it to the front end for display through the "Reply" node.

[0092] Export Report: Enter the report export process. First, classify the question through the "Question Classification" node. The questions are classified into "Query Passenger Flow Data", "Query Ticket Data", "Query Hotel Data", and "Query Scenic Spot Data". Then, simultaneously call the "Knowledge Retrieval" node to recall relevant knowledge and call the interface through the "Http Request" node to obtain data from the database. Input the data from both into the "Large Model" node. The large model performs reasoning and gives the response content. Finally, call the "Report Export Tool" to organize the data into a downloadable report, and return it to the front end for display through the "Reply" node.

[0093] The specific processing flow is as Figure 4 shown.

[0094] Start Node: Obtain the user input and save the data into the sys.query variable. Custom variables can also be defined.

[0095] Question Classification Node: By defining the classification description, the question classifier can, based on the user input, use the large model to infer the matching classification and output the classification result, providing more accurate information to the downstream nodes.

[0096] Configuration Steps: Select the input variable, which refers to the input content for classification and supports input file variables. In the customer service Q&A scenario, it is generally the question sys.query input by the user;

[0097] Select the inference model. Based on the natural language classification and inference capabilities of the large language model, choosing a suitable model will help improve the classification effect;

[0098] Write classification labels / descriptions. Multiple classifications can be added manually. By writing the keywords or description statements of the classification, the large language model can better understand the classification basis.

[0099] Select the downstream node corresponding to the classification. After the question classification node completes the classification, the subsequent process path can be selected according to the relationship between the classification and the downstream node.

[0100] Knowledge Retrieval Node: Retrieves text content related to the user's question from the knowledge base, which can be used as the context for downstream large model nodes.

[0101] Configuration Process: Select the query variable. The query variable usually represents the user's input question, which can be used as an input item to retrieve relevant text segments in the knowledge base. In common dialogue applications, the sys.query of the start node is generally used as the query variable, and the maximum query content that the knowledge base can accept is 200 characters;

[0102] Select the knowledge base to be queried. The available knowledge bases need to be pre-created in the knowledge base;

[0103] Specify the recall mode: Connect and configure downstream nodes, generally large model nodes;

[0104] Large Model Node: Invokes a large language model to process the information (natural language, uploaded files or pictures) entered by the user at the "Start" node and gives effective response information.

[0105] Configuration Steps: Select a model, which supports globally mainstream models, including the GPT series of OpenAI, the Claude series of Anthropic, the Gemini series of Google, etc. Selecting a model depends on factors such as its reasoning ability, cost, response speed, context window, etc. It is necessary to select a suitable model according to the scenario requirements and task types.

[0106] Configure model parameters. Model parameters are used to control the generation results of the model, such as temperature, TopP, maximum tokens, reply format, etc. For the convenience of selection, the system provides 3 sets of preset parameters: creative, balanced, and precise. If you are not familiar with the above parameters, select the default settings. If you want the application to have image analysis capabilities, you can select a model with visual capabilities.

[0107] Fill in the context (optional). The context can be understood as the background information provided to the large model, and is often used to fill in the output variables of knowledge retrieval.

[0108] Write prompts. The large model node provides an easy-to-use prompt arrangement page. Selecting a chat model or a completion model will display different prompt arrangement structures. If you select a chat model (Chat model), you can customize the three parts of the system prompt (SYSTEM) / user (USER) / assistant (ASSISTANT).

[0109] HTTP Request Node: Allows sending server requests via the HTTP protocol, applicable to scenarios such as obtaining external data, webhooks, generating images, downloading files, etc. It can send customized HTTP requests to specified network addresses to achieve interconnection and interoperability with various external services.

[0110] This node supports common HTTP request methods:

[0111] GET: Used to request the server to send a certain resource.

[0112] POST: Used to submit data to the server, usually for submitting forms or uploading files.

[0113] HEAD: Similar to a GET request, but the server does not return the resource body of the request, only the response headers.

[0114] PATCH: Used to obtain the transmission path at each node in the request-response chain.

[0115] PUT: Used to upload resources to the server, usually for updating existing resources or creating new resources.

[0116] DELETE: Used to request the server to delete a specified resource.

[0117] It is possible to configure the HTTP request including URL, request headers, query parameters, request body content, and authentication information, etc.

[0118] Tool Node: The "Tool" node can provide strong third-party capability support for the workflow.

[0119] Configuration steps: Authorize the tool / Create a custom tool / Publish the workflow as a tool;

[0120] Configure the tool inputs and parameters.

[0121] Direct Reply Node: Defines the reply content in a process.

[0122] It is possible to freely define the reply format in a text editor, including customizing a fixed text content, using the output variables in the previous steps as the reply content, or combining the custom text with variables for reply.

[0123] Nodes can be added at any time to stream the content to the conversation reply, supporting the what-you-see-is-what-you-get configuration mode and supporting mixed text and graphics, such as:

[0124] Output the reply content of the large model node;

[0125] Output the generated image;

[0126] Output plain text.

[0127] Example 3

[0128] The third embodiment of the present invention, an application method for realizing a travel-related enterprise assistant based on a large model, includes:

[0129] The test scenario is the intelligent customer service system of a large enterprise, which is used to handle technical support requests from customers. The front-end interface of the system supports natural language input, file upload (such as log files), and picture upload (such as device failure photos). The test data is sourced from the enterprise's historical work order records, including 1,000 real user requests, covering three categories: simple queries, complex technical problems, and cross-domain problems. The static database is the enterprise's internal knowledge base, containing 100,000 technical documents; the dynamic database is the real-time updated work order processing records, containing 50,000 pieces of data.

[0130] Implementation process

[0131] (1) Data input and status capture:

[0132] The user inputs a question through the front-end interface, and the system captures the user's in-depth information in real time, including semantic complexity (calculated through dependency syntactic analysis), interface dwell time (the time from page loading to submission), and secondary inquiry rate (the number of times the user inquires about the initial answer). The input data and status are saved in custom variables for subsequent processing.

[0133] (2) Static data retrieval:

[0134] When the input data stream exceeds the threshold (500 requests per second), the system activates the fast sorting channel. Key features (such as data fluctuation patterns, time sensitivity) are extracted through FPGA hardware acceleration technology, and a three-layer filtering screen is used for rough classification: the first layer screens out obviously irrelevant data (such as advertisements or invalid inputs), the second layer pre-groups by domain tags (such as network, hardware, software), and the third layer performs priority sorting based on historical pattern matching.

[0135] (3) Dynamic reasoning and result generation:

[0136] Configure the parameters of the large language reasoning model according to the user's in-depth information. When the semantic complexity ≥ 5 layers or the keyword density < 30%, activate the three-level retrieval mechanism: the first layer uses BERT semantic embedding matching, the middle layer uses an improved TF-IDF weighted algorithm (α = 0.65), and the bottom layer triggers cross-domain knowledge graph association. The dynamically output results are combined with the static retrieval results to form the final application output results.

[0137] (4) Output method selection and display:

[0138] The question classifier selects the output method according to the user's historical preferences and question types. For example, chart output is used for technical parameter comparison, text output is used for troubleshooting steps, and report export is used for statistical analysis results. The final result is displayed through the front-end interface.

[0139] (5)Dynamic adjustment of model parameters:

[0140] When the user's stay time > 8 seconds and the secondary questioning rate > 40%, the system starts the reinforcement learning mechanism, dynamically adjusts the model parameters through the time decay factor, and ensures the continuous optimization of the model.

[0141] Six groups of experiments were conducted in this embodiment, namely performance data for simple query scenarios, performance data for complex technical scenarios, performance data for cross-domain question scenarios, performance data for high-concurrency scenarios, performance data for dynamic parameter adjustment tests, and performance data for traditional system comparisons.

[0142] Table 1: Performance data for simple query scenarios

[0143]

[0144] Table 2: Performance data for complex technical scenarios

[0145]

[0146] Table 3: Performance data for cross-domain question scenarios

[0147]

[0148] Table 4: Performance data for high-concurrency scenarios

[0149]

[0150] Table 5: Performance data for dynamic parameter adjustment tests

[0151]

[0152] Table 6: Performance data for traditional system comparisons

[0153]

[0154] 1. Performance comparison and analysis

[0155] It can be seen from the tabular data that the system of the present invention is superior to the traditional system in all key indicators:

[0156] Retrieval time: In the high-concurrency scenario, the retrieval time of the system of the present invention is 200 milliseconds, while that of the traditional system is 600 milliseconds, and the efficiency is improved by 67%.

[0157] Accuracy: For complex technical problems, the accuracy rate of the system of the present invention reaches 89%, while that of the traditional system is only 75%, with a 14 - percentage - point increase.

[0158] User satisfaction: In the cross - domain problem scenario, the user satisfaction of the present invention is 86%, while that of the traditional system is 70%, with a 16 - percentage - point increase.

[0159] 2. Technical Advantage Analysis

[0160] Three - level retrieval mechanism: Through the association of BERT semantic embedding matching and cross - domain knowledge graphs, the system can handle complex semantic problems (such as nested compound sentences), and the accuracy rate is significantly improved.

[0161] Dynamic parameter adjustment: Based on the reinforcement learning mechanism of user stay duration and secondary questioning rate, the model can be adaptively optimized, and the user satisfaction continues to increase.

[0162] Output - mode adaptation: The problem classifier dynamically selects the output form (text, chart, report) according to user preferences, and the information transmission efficiency is increased by 40%.

[0163] 3. Advantage Summary

[0164] High efficiency: Hardware acceleration and multi - layer filtering screens enable the system to maintain millisecond - level response in high - concurrency scenarios.

[0165] Precision: The three - level retrieval mechanism and dynamic parameter adjustment ensure the accuracy and stability of complex problem processing.

[0166] User experience: Output - mode adaptation and dynamic optimization mechanisms significantly improve user satisfaction and task completion rate.

[0167] Conclusion: Through innovative workflow orchestration technology and dynamic feedback mechanisms, the present invention shows significant advantages in retrieval efficiency, accuracy, and user experience, and has clear technological progressiveness and industrial application value.

[0168] Example 4

[0169] Refer to Figure 4 , which is the fourth embodiment of the present invention. Different from the first three embodiments, the application method of the travel - related enterprise assistant based on the large model includes:

[0170] Next, refer to Figure 4, which shows a schematic structural diagram of an electronic device 300 suitable for implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 2 The enterprise assistant architecture shown is only an example and should not impose any limitations on the functions and usage scopes of the embodiments of the present invention.

[0171] As Figure 4 shown, the electronic device 300 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 301, which may perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 302 or the program loaded from the storage device 308 into the random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The processing device 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. The input / output (I / O) interface 305 is also connected to the bus 304.

[0172] Generally, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 308 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 309. The communication device 309 may allow the electronic device 300 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 4 the electronic device 300 with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. More or fewer devices may be alternatively implemented or had. Figure 4 Each block shown in

[0173] Further, the storage medium according to the embodiments of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. Application method for realizing a travel enterprise assistant based on a large model, characterized in that , The method includes: Obtaining the user's input data and user depth information through the front-end interface, saving the user depth information into a custom variable, where the input data includes natural language, files, and pictures, and the user depth information includes the semantic complexity input by the user, the user interface residence time, and the page secondary inquiry rate; Retrieving the static output result from the static database based on the input data according to the workflow orchestration technology; Configuring the model parameters according to the custom variable to establish a large language inference model; Establishing the logical address mapping relationship between the input data and the content of the dynamic database according to the large language inference model, and outputting the dynamic output result; Combining the analysis of the dynamic output result and the static output result to form an application output result; Establishing a problem classifier, and selecting the output method according to the classification of the problem classifier for the application output result. The problem classifier refers to the user's selection of the output method for the application output result, and the output method selection includes text output, generating charts, and exporting reports; Outputting the content of the application output result in the classification method output by the problem classifier and displaying it on the front-end interface; The obtaining of the user depth information through the front-end interface includes: Obtaining the semantic complexity, user residence time, and secondary inquiry rate when the user inputs at the front end; The obtaining of the semantic complexity when the user inputs at the front end includes: Judging the dependency syntactic analysis depth and keyword density in the semantic complexity. When the dependency syntactic analysis depth exceeds the threshold or the keyword density is lower than the threshold, the system starts a three-level retrieval mechanism. The first layer uses BERT-based semantic embedding matching, the middle layer uses an improved TF-IDF weighted algorithm, and the bottom layer triggers the cross-domain knowledge graph association. After the user's retrieval, the model parameters are dynamically adjusted through the user residence time and secondary inquiry rate to form a reinforcement learning mechanism with a time decay factor; The dynamically adjusting the model parameters through the user residence time and secondary inquiry rate includes that when the user residence time exceeds the residence time threshold of 8s and the secondary inquiry rate exceeds the secondary inquiry rate threshold of 40%, a reinforcement learning mechanism with a time decay factor is formed to dynamically reduce the model parameters in the large language inference model.

2. The application method for implementing a travel enterprise assistant based on a large model according to claim 1, characterized in that, The system starts a three-level retrieval mechanism when the dependency syntactic analysis depth exceeds the threshold or the keyword density is lower than the threshold, including: The first layer uses BERT-based semantic embedding matching, the middle layer uses an improved TF-IDF weighted algorithm, introducing the word vector cosine similarity correction factor α = 0.65, and the bottom layer triggers the cross-domain knowledge graph association; The dependency syntactic analysis depth exceeds the threshold or the keyword density is lower than the threshold includes: When the dependency syntactic analysis depth ≥ 5 layers or the keyword density < 30%, the retrieval mechanism is triggered.

3. The application method for realizing a travel enterprise assistant based on a large model according to claim 1, wherein, Retrieving the static output result from the static database based on the input data according to the workflow orchestration technology includes: When the application module receives a data stream exceeding the input data threshold, automatically start the fast sorting channel, where the data stream refers to the number of user access requests; Rapidly extracting key features according to the hardware acceleration technology, where the key features include data fluctuation patterns and time sensitivity; Perform rough classification according to a multi-layer filtering screen of key features. The first layer of the rough classification screens out obviously irrelevant data, the second layer pre-groups by domain tags, and the third layer performs priority sorting based on historical pattern matching.

4. The application method for implementing a travel enterprise assistant based on a large model according to claim 1, characterized in that, The classification method of outputting the application output result through a problem classifier is used for content output, and the display on the front-end interface includes: The content output includes reply text, generating charts, and exporting reports, and the content output method is freely selected by the user during the content input process; The content output method includes selecting the application output result formed by combining the dynamic output result and the static output result as the content output.

5. The application method for implementing a travel enterprise assistant based on a large model according to claim 4, wherein, The formation process of the reply text includes: The user inputs data, and the application interface obtains the user input through the sys.query variable. After obtaining the application output result, use the problem classifier to determine whether the content output method is reply text, generating charts, or exporting reports; if it is determined to display text, directly reply the text to the front-end interface; if it is determined to generate charts or export reports, stop the output and return to the problem classification stage.

6. The application method for implementing a travel enterprise assistant based on a large model according to claim 4, wherein, The formation process of generating charts includes: The user inputs data, and the application interface obtains the user input through the sys.query variable. After obtaining the application output result, use the problem classifier to determine whether the content output method is reply text, generating charts, or exporting reports; if it is determined to generate charts, obtain the x-axis data type and y-axis data type of the chart from sys.query through the parameter extractor, input the application output result into the x-axis and y-axis, generate the x and y-axis data of the chart, and call the bar chart tool node to generate the chart to the front-end interface; if it is determined to be reply text or export reports, stop the output and return to the problem classification stage.

7. The application method for implementing a travel enterprise assistant based on a large model according to claim 4, characterized in that, The formation process of exporting reports includes: The user inputs data, and the application interface obtains the user input through the sys.query variable. After obtaining the application output result, use the problem classifier to determine whether the content output method is reply text, generating charts, or exporting reports; if it is determined to export reports, obtain the horizontal data type and vertical data type of the report from sys.query through the parameter extractor, call the report tool to generate a report, input the application output result into the horizontal data type and vertical data type, and generate the report to be displayed on the front-end interface; if it is determined to be reply text or generate charts, stop the output and return to the problem classification stage.

8. An electronic device, comprising: One or more processors; A storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-7.

9. A computer-readable storage medium having executable instructions stored thereon, which when executed by a processor cause the processor to implement the method according to any one of claims 1-7.

Images