Large model arbitration-based degradable dialogue analysis method, system and electronic device

Abstract

The application provides a degradable dialogue analysis method and system based on a large model arbitration, and an electronic device, and relates to the technical field of data analysis. The method comprises the following steps: storing a plurality of BI components preset in a component registration center; obtaining a user natural language query input, and screening a BI component most matched with the user natural language query input from a plurality of candidate BI components through a large language model; performing parameter verification on the BI component most matched with the user natural language query input, performing component rendering on the BI component most matched with the user natural language query input and passing the parameter verification, and obtaining a query result; and if there is no BI component most matched or the parameter verification fails, degrading to generate an SQL query according to an NL2SQL technology and obtaining a corresponding query result. The application preferentially reuses verified inventory BI components, ensures the accuracy of the analysis result, and the response speed is much faster than that of the form of the real-time NL2SQL generated query result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data analysis, and particularly relates to a degradable dialogue analysis method and system based on large model arbitration and an electronic device. BACKGROUND

[0002] With the explosive development of large language model (LLM) technology, dialogue-based analysis based on structured data has become the core form of intelligent data analysis, and is experiencing a technical evolution from NL2SQL generation to multi-path hybrid response. The current mainstream solutions include:

[0003] NL2SQL / NL2DSL generation route: through a large model, natural language is directly converted into SQL (database language) or DSL (domain-specific language), and the accuracy of the answer is improved by relying on intent clarification and ICL optimization.

[0004] Semantic layer unification route: adopt to construct index and label semantic layer, realize natural language to business index mapping, and avoid the instability of bottom SQL generation.

[0005] Multi-agent collaboration route: adopt task division and resolution agent to reduce the complexity of end-to-end generation.

[0006] The mainstream solutions are constantly iterated and upgraded, but the technical path still focuses on language parsing optimization. There are the following limitations:

[0007] (1) Insufficient fuzzy intent query capability.

[0008] When the user's question does not explicitly analyze the dimension and calculation index (such as "How is the company's efficiency this year?"), the existing technology is difficult to accurately map to the specific fields in the database, and the generated query has a high probability of missing key business indicators, or even cannot return the result.

[0009] (2) Low utilization rate of enterprise-constructed BI assets (BI components).

[0010] The BI system constructed by the enterprise usually contains a large number of pre-defined BI components, which have fixed analysis logic and verified value. However, the existing dialogue-based analysis generally generates visual charts based on data sets and analysis models, ignoring the reuse of existing components.

[0011] (3) Insufficient robustness of single NL2SQL or NL2DSL path.

[0012] When the user query input cannot match the database field, the existing system lacks a degradation mechanism, usually directly reporting an error or requiring the user to re-input.

[0013] (4) Low accuracy of complex parameter parsing.

[0014] For the query conditions involving the use of ID, encoding, etc., and complex conditions such as multiple selection and single selection, the query accuracy is low. SUMMARY

[0015] To overcome the deficiencies of the prior art, the present application provides a degradable dialogue analysis method and system based on large model arbitration, which preferentially reuses verified inventory BI components to ensure the accuracy of the analysis results and the response speed far exceeds the form of NL2SQL generated query results.

[0016] To achieve the above object, one or more embodiments of the present application provide the following technical solutions:

[0017] The first aspect of the present application provides a degradable dialogue analysis method based on large model arbitration.

[0018] The degradable dialogue analysis method based on large model arbitration comprises the following steps:

[0019] The plurality of BI components preset are stored to the component registration center;

[0020] The user natural language query input is obtained, and a plurality of candidate BI components most relevant to the user natural language input are found from the component registration center;

[0021] The BI component most matched to the user natural language query input is screened from the plurality of candidate BI components through a large language model;

[0022] The BI component most matched to the user natural language query input is subjected to parameter analysis and parameter verification, and the BI component most matched to the user natural language query input that passes the parameter verification is subjected to component rendering to obtain a query result;

[0023] If there is no BI component most matched or the parameter verification fails, the system is degraded to generate a SQL query according to the NL2SQL technology and obtains a corresponding query result;

[0024] The obtained query result is presented to the user.

[0025] The second aspect of the present application provides a degradable dialogue analysis system based on large model arbitration.

[0026] The degradable dialogue analysis system based on large model arbitration comprises:

[0027] The BI component storage module is configured to store the plurality of BI components preset to the component registration center;

[0028] The one-time screening module is configured to obtain the user natural language query input, and find a plurality of candidate BI components most relevant to the user natural language input from the component registration center;

[0029] The secondary screening module is configured to screen, by the large language model, a BI component that is most matched with the user natural language query input from the plurality of candidate BI components;

[0030] The parameter verification module is configured to perform parameter analysis on the BI component that is most matched with the user natural language query input, and perform parameter verification, and perform component rendering on the BI component that is most matched with the user natural language query input and passes the parameter verification to obtain a query result;

[0031] The degradation processing module is configured to, if there is no BI component that is most matched or the parameter verification does not pass, degrade to generate a SQL query according to an NL2SQL technology and obtain a corresponding query result;

[0032] The presentation module is configured to present the obtained query result to the user.

[0033] The third aspect of the present application provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to realize the steps in the degradable dialogue analysis method based on large model arbitration as described in the first aspect of the present application.

[0034] The above one or more technical solutions have the following beneficial effects:

[0035] The present application provides a degradable dialogue analysis method, system and electronic device based on large model arbitration, which sets two logical execution paths: preferentially reuse verified inventory BI components, and degrade to execute NL2SQL technology to generate SQL queries. The two logical execution paths not only ensure the accuracy and timeliness of the analysis results, improve the reuse rate of BI assets, but also provide a way to generate SQL queries with higher priority than traditional NL2SQL technology, improving the robustness of existing technology relying on single NL2SQL technology to generate SQL queries.

[0036] The unique LLM arbitration engine of the present application can intelligently judge the matching degree of user intent and existing BI components, dynamically select the optimal response path, and make decisions on reuse or degradation, greatly improving the system's ability to handle complex and ambiguous queries.

[0037] The present application can "reverse feed" valuable new analysis charts generated successfully through the degradation path into new BI components, forming a closed loop of continuous learning and self-improvement.

[0038] In the process of screening the BI component most matched with the user natural language query input, the BI components of the component registration center are pre-filtered based on the keyword matching mode first, vector similarity calculation is performed on the BI components obtained by pre-filtering and the query vector, and a plurality of candidate BI components are obtained; then the BI component most matched with the user natural language query input is screened from the plurality of candidate BI components through the mode of calculating the confidence score, and the matching speed is improved.

[0039] The user natural language query input is converted into a completely structured parameter object, and then parameter verification with the most matched BI component is realized, so that the problem of insufficient query capability caused by fuzzy intention can be better solved. The completely structured parameter object is taken as a reference parameter, and whether the BI component most matched with the user natural language query input after analysis covers all reference parameters is compared, if all reference parameters are covered, the parameter verification is passed, if at least one reference parameter is not covered, the parameter verification is not passed. The BI component most matched with the user natural language query input and passing the parameter verification is rendered as a component, and a query result is obtained, which is pushed and presented to the user.

[0040] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of the specification, are included to provide a further understanding of the application and are incorporated herein by reference. The illustrations are shown for the purpose of enabling those skilled in the art to implement the application and are not intended to limit the scope of the application.

[0042] Figure 1 The method flowchart of the embodiment is shown in Figure 1.

[0043] Figure 2 The method block diagram of the embodiment is shown in Figure 2. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0045] It should be noted that the terms used herein are only for the purpose of describing the specific embodiments and are not intended to limit the exemplary embodiments according to the application.

[0046] The embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0047] Term explanation:

[0048] LLM: large language model;

[0049] NL2SQL: converting natural language to database language;

[0050] NL2DSL: converting natural language to domain-specific language;

[0051] SQL: database language;

[0052] DSL: domain-specific language;

[0053] ICL: context learning tuning;

[0054] BI component: visualization chart and associated metadata configured in existing BI system.

[0055] Embodiment I

[0056] Although the current chat analysis (ChatBI) field has made progress in semantic parsing, knowledge base construction, query acceleration, etc., it still does not systematically solve the problem of "fuzzy intent-prepared assets-degraded execution" coordination. In order to solve this problem, this embodiment proposes a degradable dialogue analysis method based on large model arbitration, which realizes dynamic semantic recall of BI prepared assets (BI components) and LLM driven double path arbitration mechanism, realizes the closed loop reuse of enterprise analysis assets, and makes up for the insufficient robustness of single NL2SQL or NL2DSL query path.

[0057] As shown in Figure 1 , the degradable dialogue analysis method based on large model arbitration includes the following steps:

[0058] Store a plurality of preset BI components to the component registration center;

[0059] Obtain user natural language query input, and find a plurality of candidate BI components most relevant to the user natural language input from the component registration center;

[0060] Filter out the BI component most matched to the user natural language query input from the plurality of candidate BI components through a large language model;

[0061] Perform parameter analysis on the BI component most matched to the user natural language query input, and perform parameter verification. The BI component most matched to the user natural language query input that passes the parameter verification is rendered to obtain a query result;

[0062] If there is no BI component most matched or the parameter verification does not pass, degrade to generate a SQL query according to the NL2SQL technology, and obtain the corresponding query result;

[0063] Present the obtained query result to the user.

[0064] NL2SQL stands for Natural Language to SQL. It enables real-time data queries between queries and relational databases, allowing users to manipulate data flexibly and quickly using natural language. Machines can understand user descriptions and retrieve answers from database tables. NL2SQL has significant practical value, improving the way users interact with databases.

[0065] Business Intelligence (BI) refers to the use of modern data warehousing technology, online analytical processing technology, data mining, and data visualization technology to analyze data and realize business value. It effectively integrates existing data within an enterprise, quickly and accurately provides reports and decision-making support, and helps enterprises make informed business decisions. It generally includes components such as data warehousing, report querying, data analysis, data mining, and data visualization.

[0066] This embodiment adopts the "recall-arbitration-dual-path execution" logic, and the flowchart is as follows. Figure 2 As shown, it includes six core steps: component registration center, semantic recall, LLM arbitration engine, component rendering, parameter parsing and validation, LN2SQL degradation, and closed-loop feedback. The specific execution process includes:

[0067] First, user natural language query input is processed. Based on the user natural language query input, semantic recall is performed on BI components in the component registry. At the same time, this step also performs semantic recall on other knowledge such as model table structure.

[0068] During the semantic recall process, multiple candidate BI components most relevant to the user's natural language query input are identified, packaged, and sent to the LLM arbitration stage.

[0069] Through LLM arbitration, a two-path decision is made to determine whether to reuse BI components or downgrade them.

[0070] When reusing BI components, it is necessary to use the component parameter parsing engine to perform parameter validation between the user's natural language query input and the best matching BI component. Then, the BI component that passes the parameter validation and best matches the user's natural language query input will be rendered to obtain the query results.

[0071] If no best-matching BI component is found or parameter validation fails, the process is downgraded to generating an SQL query using NL2SQL technology and obtaining the corresponding query results. Finally, the query results are presented.

[0072] The method of the embodiment can be applied in a system for realizing data analysis through natural language interaction, deeply integrating a large language model (LLM), enhanced retrieval generation (RAG), and reuse of pre-set BI assets, and solving the problem of accurate response to fuzzy intent queries in a conversational analysis scenario.

[0073] Next, the method of the embodiment will be explained in detail.

[0074] I. Component registration center

[0075] The component registration center is a management place of reusable BI components, responsible for metadata definition and storage of BI components, and dynamic semantic index establishment.

[0076] In the embodiment, a plurality of preset BI components are stored in the component registration center, specifically including:

[0077] Metadata definition of the BI component: each BI component is abstracted as an entity containing semantic information, and the metadata includes the component ID, component name, business description, synonyms or alias, analysis dimension, analysis indicator, parameterized variable, and chart type of the BI component;

[0078] Dynamic semantic index creation for the BI component: core text information is extracted from the metadata of the BI component, including the component name, business description, analysis dimension, and analysis indicator, and based on the core text information, a semantic vector is generated by calling a text embedding model; the component ID of the BI component and the corresponding semantic vector are stored in a vector database.

[0079] Specifically:

[0080] 1. Component metadata structure definition: each BI component is abstracted as an entity containing rich semantic information.

[0081] The metadata includes:

[0082] wedget_id: component ID, used for uniquely identifying the BI component.

[0083] wedget_name: component name, such as 30-day sales trend.

[0084] widget_description: business description, such as "show the changes of the company's daily total sales to monitor short-term performance fluctuations".

[0085] widget_synonyms: synonyms or alias, such as "sales board" and "performance dashboard".

[0086] widget_dimensions: Analysis dimension list, such as [“Date”, “Region”].

[0087] widget_measures: Analysis measure list, such as [“Sales”, “Profit”].

[0088] widget_params: Parameterized variable list, defining the interaction ability of the component. Each parameter contains parameter name param_name (such as time range time_range), parameter type param_type (such as date type Date), parameter default value default_value, etc.

[0089] chart_type: Chart type, such as LineChart (line chart), BarChart (bar chart).

[0090] 2. Dynamic semantic index creation and maintenance:

[0091] Vectorization: When a BI component is registered or updated, the system automatically extracts the core text information (name, description, dimension, measure, etc.) in its metadata, calls a text embedding model (such as BGE-M3, Sentence-BERT) to generate a high-dimensional semantic vector.

[0092] Index storage: Store the component ID and its corresponding high-dimensional semantic vector in a special vector database.

[0093] II. Semantic recall

[0094] The semantic recall step is responsible for quickly and accurately retrieving the most relevant candidate BI components from the component registration center according to the user's natural language query input. In addition, it also includes retrieving model structure knowledge, data field knowledge, and enterprise private domain knowledge from the knowledge base.

[0095] In this step, the user's natural language query input is obtained, and the most relevant candidate BI components are found from the component registration center. Specifically, it includes:

[0096] Vectorize the user's natural language query input to get the query vector;

[0097] Get the keywords from the user's natural language query input, and use keyword matching to pre-filter the BI components in the component registration center;

[0098] Calculate the vector similarity between the pre-filtered BI components and the query vector to find the Top-K candidate BI components that are closest in semantics to the user's natural language query input.

[0099] Specifically comprising:

[0100] 1. Query vectorization:

[0101] The natural language input by the user (Query) is converted into a query vector using the same embedding model as when creating the dynamic semantic index of the BI component.

[0102] 2. Metadata pre-filtering:

[0103] Before performing vector search, a lightweight pre-filtering step is performed. For example, if the user query is "show the sales of mobile phone categories in the East China region", the system will first filter out a subset of components whose metadata description, dimension or indicator field contains "sales", "category" and "region" through keyword matching from the full component library. This can reduce the search range by more than 90%, greatly improving efficiency and accuracy.

[0104] 3. Vector similarity search:

[0105] On the pre-filtered component subset, vector similarity calculation (such as cosine similarity) is performed to find the Top-K candidate components (e.g. K=5) that are closest in semantics to the user query vector.

[0106] 4. Output:

[0107] The K candidate component IDs and their complete metadata are packaged and sent to the LLM arbitration engine.

[0108] Three, LLM arbitration engine

[0109] This is the decision-making brain of the system, responsible for making the best choice between the "reuse" and "degrade" paths. Its core is a decision-making mechanism that combines quantitative evaluation and Few-Shot Prompting (small sample prompting).

[0110] 1. Quantitative confidence evaluation and dynamic prompt word construction:

[0111] In order to guide the LLM to make stable, accurate and controllable judgments, the system dynamically constructs a prompt word containing quantitative scoring rules and context examples (Few-Shot).

[0112] Further, in this embodiment, the large language model selects the BI component that best matches the user's natural language query input from multiple candidate BI components by calculating a confidence score, and the specific calculation formula for the confidence score is:

[0113] Score = a×F_param + b×F_measure + c×F_dim + d×F_name;

[0114] Wherein, Score is the score of the confidence score; F_param is the parameter matching score; F_measure is the analysis index matching score; F_dim is the analysis dimension matching score; F_name is the component name matching score; a, b, c, d are weights, a>b>c>d, a+b+c+d=1.

[0115] Wherein:

[0116] The parameter matching score F_param is calculated as: the number of parameterized variables in the user natural language query input that can match the BI component / the total number of parameterized variables of the BI component;

[0117] The analysis index matching score F_measure is used to measure the text similarity of the analysis index in the user natural language query input and the analysis index of the BI component;

[0118] The analysis dimension matching score F_dim is used to measure the text similarity of the analysis dimension in the user natural language query input and the analysis dimension of the BI component;

[0119] The component name matching score F_name is used to measure the text similarity of the user natural language query input and the component name of the BI component.

[0120] Further, when the score of the confidence score is less than a set score threshold, it is determined that there is no BI component in the plurality of candidate BI components that best matches the user natural language query input;

[0121] When the score of the confidence score is greater than the set score threshold, it is determined that there is a BI component in the plurality of candidate BI components that best matches the user natural language query input, and the candidate BI component with the highest confidence score is selected as the best matching BI component.

[0122] In this embodiment, the set score threshold is 0.7; a, b, c, d are set to 0.4, 0.3, 0.2 and 0.1 respectively.

[0123] Further, the prompt word template is as follows:

[0124] # Role

[0125] You are a senior BI analysis expert, and your task is to select the best matching component for reuse according to the user's question and a series of candidate BI components, or to determine that all components are not applicable and to decide to degrade.

[0126] # Decision Guide

[0127] You need to calculate a confidence score (between 0 and 1) for each candidate component and make a final decision based on the score.

[0128] # Confidence Score Calculation Formula

[0129] Score = 0.4×F_param + 0.3×F_measure + 0.2×F_dim + 0.1×F_name

[0130] - F_param: Parameter Matching Score. Calculation method: (Number of parameters in the question that can be matched) / (Total number of component parameters).

[0131] - F_measure: Indicator Matching Score. Measures the text similarity of the analysis indicators in the question and the component indicators (using cosine similarity).

[0132] - F_dim: Dimension Matching Score. Measures the text similarity of the analysis dimensions in the question and the component dimensions.

[0133] - F_name: Name Matching Score. Measures the text similarity of the question and the component name.

[0134] # Decision Rules

[0135] - If the highest score Score>0.7, consider it a high match, decide to "reuse", and return the component ID.

[0136] - If the highest score Score<=0.7, consider it a low match, decide to "downgrade".

[0137] # Current Task

[0138] - User Question: "{User Original Query}"

[0139] - Candidate Component List:

[0140] {Component1_ID}: Name={Name1}, Dimension={Dimension1}, Indicator={Indicator1}, Parameter={Parameter1; Parameter2, Parameter3}

[0141] {Component2_ID}: Name={Name2}, Dimension={Dimension2}, Indicator={Indicator2}, Parameter={Parameter2} ...

[0142] - Please output your thinking process and final decision.

[0143] # Output Format

[0144] <thinking>

[0145] [Your thought process]

[0146] < / thinking>

[0147] <decision>

[0148] Component ID or "Degradation"

[0149] < / decision>

[0150] 2. Call the large LLM model and parse the output:

[0151] The system sends a pre-filled Prompt containing the user's question and candidate part information to a large model (such as DeepSeek, GLM, Qwen, etc.). The model returns a structured text (XML or JSON) containing the thought process and final decision. The system parses the result; if it determines the part ID, it calls the part rendering engine; if it determines the downgrade, it triggers the NL2SQL downgrade unit.

[0152] IV. Component Parameter Parsing Engine

[0153] Keywords extracted from the user's natural language are precisely mapped to standardized parameter values ​​required by the BI component, and then verified. Any parameters that do not meet the requirements are rejected outright.

[0154] Furthermore, the BI component that best matches the user's natural language query input will undergo parameter parsing and validation, specifically:

[0155] Entity extraction is performed on the user's natural language query input to obtain key entities;

[0156] Standardize key entities;

[0157] Check if there are any other default parameters in the user's natural language query input, and if so, complete them;

[0158] The standardized key entities are assembled with other default parameters to obtain a fully structured parameter object;

[0159] The fully structured parameter object is used as the baseline parameter. The BI component that best matches the user's natural language query input after parsing is compared to see if it covers all the baseline parameters. If all parameters are covered, the parameter validation passes. If at least one parameter is not covered, the validation fails.

[0160] Its core is a parameter-standardized intelligent agent. This "intelligent agent" can be implemented by a standalone LLM fine-tuned with specific instructions, or by a hybrid system combining named entity recognition (NER), a rule engine, and a dictionary.

[0161] The workflow example is as follows:

[0162] 1. User question: "Compare the sales figures for the East China and North China regions in the first half of this year."

[0163] 2. Required parameters for target component: region parameter region: List[str], time range time_range: [start_date, end_date].

[0164] 3. Processing steps:

[0165] (1) Entity extraction (NER):

[0166] The agent extracts key entities from the question: "East China", "North China" (regions), and "this year's first half" (time).

[0167] (2) Entity linking and standardization:

[0168] A. Query the internal geographic knowledge base or synonym dictionary for "East China" and "North China" to map them to standard region unique identifiers region ID: ["east_china", "north_china"].

[0169] B. According to the current date (assuming 2025-07-26), call the time resolution module to resolve the relative time "this year's first half" to an absolute date range: start time start_date: "2025-01-01", end time end_date: "2025-06-30".

[0170] (3) Default value injection:

[0171] Check if there are any other necessary parameters in the Query that are not mentioned. For example, if the default parameter currency is also required, the agent can automatically fill in the default value "CNY" based on the user profile or system configuration.

[0172] (4) Assemble the final parameter:

[0173] Get the fully structured parameter object: { "region": ["east_china", "north_china"], "time_range": ["2025-01-01", "2025-06-30"], "currency": "CNY"}.

[0174] (5) Parameter verification confirmation:

[0175] Check if the obtained parameter object conditions are complete and if there are any mandatory items that are empty. If any of these conditions are not met, further component rendering is not performed, and the NL2SQL logic is degraded.

[0176] Five, component rendering engine

[0177] Responsible for rendering the parameters and final chart when the above decision is "reuse".

[0178] Using the component ID and the fully structured parameter object, call the API (Application Programming Interface) of the BI system, execute the data query and render the final visualization chart, and present it to the user.

[0179] Six, NL2SQL degradation and closed-loop counter-benefits

[0180] 1. Degradation execution:

[0181] When the arbitration engine decides to degrade, the system passes the user's original Query to this unit. This unit uses traditional NL2SQL technology to generate a SQL query, execute it, and render a new, newly generated chart. This part can use existing technology and will not be described here.

[0182] 2. Closed-loop counter-benefits mechanism:

[0183] Next to the newly generated chart, there is a "adopt as standard component" or "one-click registration" button. When the user thinks that this chart has reuse value, they can click this button. The system will automatically trigger a workflow to analyze the query logic of this chart (dimensions, measures, filter conditions) and guide the user to supplement the necessary metadata (such as title, description), and finally automatically register it in the component registration center, completing the learning closed loop.

[0184] This newly registered component can be directly reused by other users through semantic recall in the future.

[0185] This embodiment has the following advantages:

[0186] (1) High fidelity and high efficiency: preferentially reuse verified stock BI components to ensure the accuracy of business logic and consistency of analysis results, while responding much faster than real-time NL2SQL generation.

[0187] (2) Excellent robustness and intelligence: the unique LLM arbitration engine can intelligently judge the matching degree of user intent and existing BI components, dynamically select the optimal response path (reuse or degradation), and greatly improve the system's ability to handle complex and ambiguous queries.

[0188] (3) Maximize enterprise asset value: activate the enterprise's dormant BI assets and convert them into dynamically callable, semantically searchable service resources, maximizing information system investment returns.

[0189] (4) Closed-loop self-evolution capability: the system adopts valuable new analysis charts successfully generated through the degradation path and registers them as new BI components, forming a closed loop of continuous learning and self-improvement, making the system smarter with use.

[0190] Embodiment Two

[0191] The embodiment discloses a degradable dialogue analysis system based on large model arbitration.

[0192] The degradable dialogue analysis system based on large model arbitration comprises:

[0193] A BI component storage module is configured to store a plurality of preset BI components to a component registration center.

[0194] A first screening module is configured to obtain a user natural language query input, and find a plurality of candidate BI components most relevant to the user natural language input from the component registration center.

[0195] A second screening module is configured to screen a BI component most matched to the user natural language query input from the plurality of candidate BI components through a large language model.

[0196] A parameter verification module is configured to perform parameter analysis on the BI component most matched to the user natural language query input, and perform parameter verification, and perform component rendering on the BI component most matched to the user natural language query input passing the parameter verification to obtain a query result.

[0197] A degradation processing module is configured to degrade to generate a SQL query according to an NL2SQL technology and obtain a corresponding query result if there is no BI component most matched or the parameter verification does not pass.

[0198] A presentation module is configured to present the obtained query result to the user.

[0199] Embodiment Three

[0200] The purpose of the embodiment is to provide an electronic device.

[0201] The electronic device comprises a memory, a processor, and a program stored on the memory and executable on the processor, and the processor executes the program to implement the steps in the degradable dialogue analysis method based on large model arbitration according to Embodiment 1 of the present disclosure.

[0202] The steps and methods involved in the above embodiments two and three correspond to Embodiment One, and the specific implementation can refer to the relevant description part of Embodiment One. The term "computer readable storage medium" should be understood to include a single medium or multiple media of one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying instruction sets for execution by a processor and causing the processor to perform any method in the present application.

[0203] Those skilled in the art should understand that the modules or steps of the present application described above can be realized by general computer devices, or alternatively, they can be realized by program codes executable by the computer devices, so that they can be stored in the storage devices and executed by the computer devices, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0204] The specific embodiments of the present application described above in conjunction with the accompanying drawings are not intended to limit the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for degradable dialogue analysis based on large model arbitration, characterized in that, The method comprises the following steps: storing a plurality of preset BI components into a component registration center; obtaining a user natural language query input, and preferentially finding a plurality of candidate BI components most relevant to the user natural language input from the component registration center; screening, by a large language model, a BI component most matched to the user natural language query input from the plurality of candidate BI components; performing parameter analysis on the BI component most matched to the user natural language query input, and performing parameter checking, and performing component rendering on the BI component most matched to the user natural language query input that passes the parameter checking to obtain a query result; if there is no BI component most matched or the parameter checking does not pass, then downgrading to generating a SQL query according to an NL2SQL technology and obtaining a corresponding query result; presenting the obtained query result to the user; performing parameter analysis on the BI component most matched to the user natural language query input, and performing parameter checking, specifically: performing entity extraction on the user natural language query input to obtain key entities; standardizing the key entities; checking whether there are other default parameters in the user natural language query input, and if so, supplementing the complete parameters; assembling the standardized key entities and other default parameters to obtain a completely structured parameter object; comparing the completely structured parameter object as a benchmark parameter with the BI component most matched to the user natural language query input after analysis to determine whether all benchmark parameters are covered, if all benchmark parameters are covered, the parameter checking passes, if at least one benchmark parameter is not covered, the parameter checking does not pass; after downgrading to generating a SQL query according to an NL2SQL technology and obtaining a corresponding query result, further comprising: obtaining a query result of whether the user establishes BI component reuse; analyzing the query logic of the query result to guide the user to supplement the complete metadata; automatically registering the supplemented metadata to the component registration center as a new BI component to participate in the next round of reuse.

2. The large model arbitration based degradable dialogue analysis method of claim 1, wherein, storing a plurality of preset BI components into a component registration center, specifically comprising: metadata definition of the BI component: each BI component is abstracted as an entity containing semantic information, the metadata includes component ID, component name, business description, synonym or alias, analysis dimension, analysis index, parameterized variable and chart type of the BI component; creating a dynamic semantic index for the BI component: extracting core text information from the metadata of the BI component, the core text information includes component name, business description, analysis dimension and analysis index, generating a semantic vector based on the core text information by calling a text embedding model; storing the component ID of the BI component and the corresponding semantic vector in a vector database.

3. The large model arbitration based degradable dialogue analysis method of claim 1, wherein, obtaining a user natural language query input, and preferentially finding a plurality of candidate BI components most relevant to the user natural language input from the component registration center, specifically comprising: vectorizing the user natural language query input to obtain a query vector; obtaining keywords from the user natural language query input, and pre-filtering the BI components of the component registration center by keyword matching; The BI component obtained by pre-filtering is subjected to vector similarity calculation with the query vector to find the Top-K candidate BI components closest to the user natural language query input in semantics.

4. The large model arbitration based degradable dialogue analysis method of claim 1, wherein, The large language model filters the BI component that best matches the user natural language query input from multiple candidate BI components by calculating a confidence score, and the specific calculation formula of the confidence score is: Score = a×F_param + b×F_measure + c×F_dim + d×F_name; Wherein, Score is the score of the confidence score; F_param is the parameter matching score; F_measure is the analysis index matching score; F_dim is the analysis dimension matching score; F_name is the component name matching score; a, b, c, d are weights, a>b>c>d, a+b+c+d=1; wherein: The calculation method of the parameter matching score F_param is: the ratio of the number of parameterized variables in the user natural language query input that can match the BI component to the total number of parameterized variables of the BI component; The analysis index matching score F_measure is used to measure the text similarity of the analysis index in the user natural language query input and the analysis index of the BI component; The analysis dimension matching score F_dim is used to measure the text similarity of the analysis dimension in the user natural language query input and the analysis dimension of the BI component; The component name matching score F_name is used to measure the text similarity of the user natural language query input and the component name of the BI component.

5. The large model arbitration-based degradable dialogue analysis method of claim 4, wherein: When the score of the confidence score is less than a set score threshold, it is determined that there is no BI component that best matches the user natural language query input in the multiple candidate BI components; When the score of the confidence score is greater than the set score threshold, it is determined that there is a BI component that best matches the user natural language query input in the multiple candidate BI components, and the candidate BI component with the highest confidence score is taken as the best matching BI component.

6. The large model arbitration based degradable dialogue analysis method of claim 1, wherein, The BI component that best matches the user natural language query input is subjected to component rendering through parameter verification to obtain a query result, specifically: Obtain the component ID of the best matching BI component; Use the component ID of the best matching BI component and the fully structured parameter object to call the API of the BI system, execute data query and render the final visual chart.

7. A degradable dialogue analysis system based on large model arbitration, characterized in that, It includes: A BI component storage module configured to store a plurality of preset BI components in a component registry center; A primary screening module configured to obtain a user natural language query input and preferentially find a plurality of candidate BI components most relevant to the user natural language input from the component registry center; A secondary screening module configured to filter the BI component that best matches the user natural language query input from the multiple candidate BI components by a large language model. The parameter checking module is configured to perform parameter analysis on the BI component that is most matched with the user natural language query input and perform parameter checking, perform component rendering on the BI component that is most matched with the user natural language query input and passes the parameter checking, and obtain a query result; The degradation processing module is configured to, if there is no BI component that is most matched or the parameter checking does not pass, degrade to generate a SQL query according to an NL2SQL technology and obtain a corresponding query result; The presentation module is configured to present the obtained query result to the user; The parameter analysis and parameter checking on the BI component that is most matched with the user natural language query input are specifically as follows: performing entity extraction on the user natural language query input to obtain key entities; standardizing the key entities; checking whether there are other default parameters in the user natural language query input, and if so, supplementing the default parameters; assembling the standardized key entities and the other default parameters to obtain a completely structured parameter object; comparing the completely structured parameter object as a benchmark parameter with the BI component that is most matched with the user natural language query input to determine whether the BI component covers all the benchmark parameters, if all the benchmark parameters are covered, the parameter checking passes, and if there is at least one benchmark parameter that is not covered, the parameter checking does not pass; after degrading to generate a SQL query according to an NL2SQL technology and obtaining a corresponding query result, further comprising: obtaining a query result of whether the user establishes BI component reuse; analyzing query logic of the query result to guide the user to supplement metadata; automatically registering the supplemented metadata to a component registration center as a new BI component to participate in the next round of reuse.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized by The processor implements the steps in the method for degradable dialogue analysis based on large model arbitration according to any one of claims 1-6 when executing the program.

Images