Intelligent question answering method based on large model and knowledge retrieval and electronic device

Abstract

This application provides an intelligent question-answering method and electronic device based on a large model and knowledge retrieval, relating to the field of artificial intelligence. The method includes: acquiring a query statement and determining the user's intent based on the query statement; performing a search and matching operation in an experience-based question-answering database based on the query statement, and determining the matching result and matching degree; in response to a matching degree less than a first preset threshold, retrieving contextual information from a pre-built business database; determining at least one task flow component and its execution order using a large model; executing each task flow component in the execution order to obtain response information and visualization rendering data; fusing the response information and visualization rendering data to obtain response text; and outputting the response text and performing a visual interactive display based on the visualization rendering data. This application not only provides users with response text matching their query statements but also enhances the spatiotemporal visualization capabilities of business data containing spatial attributes.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent question-answering method and electronic device based on large models and knowledge retrieval. Background Technology

[0002] Currently, the application of general-purpose large-scale models in vertical business domains has significant limitations. A lack of specialized domain knowledge and insufficient understanding of industry terminology and business logic leads to poor question-and-answer accuracy. Furthermore, the fragmented storage of business system data makes it difficult to effectively link professional knowledge with real business data. Large-scale models often fail to accurately query data due to a lack of understanding of database table structures and business entity relationships. In addition, traditional intelligent question answering systems mostly output text results, lacking intuitive spatiotemporal visualization capabilities for business data with spatial attributes, making it difficult for users to quickly perceive the spatial distribution and dynamic changes of the data. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose an intelligent question answering method and electronic device based on large models and knowledge retrieval, so as to solve the problems of poor accuracy and lack of spatiotemporal visualization capabilities of general large models.

[0004] To achieve the above objectives, the first aspect of this application provides an intelligent question-answering method based on large models and knowledge retrieval, comprising:

[0005] Obtain the user's input query statement and determine the user's intent based on the query statement; The search and matching are performed in a preset experience question and answer database based on the query statement to determine the matching results and the matching degree of the matching results; In response to the matching degree being less than a first preset threshold, the context information corresponding to the query statement is retrieved from a pre-built business database based on the query statement and the user intent; Based on the user intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model; Each task process component is executed in the described execution order to obtain response information and visual rendering data; The response information and the visualization rendering data are fused to obtain the response text corresponding to the query statement; The response text is output and a visual interactive display is performed based on the visualization rendering data.

[0006] Based on the same inventive concept, a second aspect of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor, when executing the computer program, implements the method described in the first aspect.

[0007] As can be seen from the above, the intelligent question-answering method and electronic device based on large model and knowledge retrieval provided in this application includes: acquiring a user-inputted query statement, and determining the user's intent based on the query statement. A search and matching operation is performed in a preset experience question-answering database based on the query statement to determine the matching result and the matching degree of the matching result. The experience question-answering database can provide historical accurate experience to increase the accuracy of the large model's response. As the system's operating frequency increases, the accumulated content of the experience question-answering database gradually increases, and the accuracy of the system's response also gradually increases. In response to the matching degree being less than a first preset threshold, indicating that no matching result with a high matching degree was found in the experience question-answering database, the context information corresponding to the query statement is retrieved from a pre-constructed business database based on the query statement and the user's intent. The business database stores domain knowledge closely related to the business domain, which can be used as context information for subsequent large models to provide precisely adapted content. Based on the user's intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined through the large model, that is, the planning of the automated link is realized through the large model. By analyzing the business exploration logic, the execution order of each task process component can be determined, which helps to accurately determine the corresponding response information for the query statement. The modular design of the task process components and the clear context information enable the system to quickly adapt to different vertical industries through configuration and knowledge updates. By flexibly arranging the execution order of the task process components, it can handle various business scenarios ranging from simple queries to complex multi-step analyses. Each task process component is executed in the execution order to obtain response information and visualization rendering data. The response information and the visualization rendering data are fused to obtain the response text corresponding to the query statement. The response text is output and displayed interactively based on the visualization rendering data. This application not only provides users with response text matching the query statement, but also provides users with more intuitive interactive visualization information, improving the spatiotemporal visualization capabilities of business data with spatial attributes. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the intelligent question-answering method based on large models and knowledge retrieval, as described in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an intelligent question-answering device based on large models and knowledge retrieval, according to an embodiment of this application. Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0011] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0012] General-purpose large-scale models have significant limitations in application within vertical business domains. A lack of specialized domain knowledge and insufficient understanding of industry terminology and business logic leads to poor question-and-answer accuracy. Furthermore, the fragmented storage of business system data makes it difficult to effectively link professional knowledge with real-world business data. Large-scale models often fail to accurately query data due to a lack of understanding of database table structures and business entity relationships. In addition, traditional intelligent question answering systems primarily output text results, lacking intuitive spatiotemporal visualization capabilities for business data with spatial attributes, making it difficult for users to quickly perceive the spatial distribution and dynamic changes of the data.

[0013] While related technologies have achieved a preliminary integration of large language models and map visualization, many problems still exist, making it difficult to meet the precision and integration needs of various professional business fields: (1) The data association and query capabilities are weak. The relevant technologies do not involve the parsing and association of professional business database table structures. They can only process preset structured data or simple location information. They cannot realize the accurate query and call of large models to dispersed business data, and it is difficult to support deep data interaction in complex business scenarios.

[0014] (2) Insufficient collaboration between question answering and visualization. Although the relevant technologies have the ability to interact, they focus on scene creation and basic operations, and have not achieved full-link collaboration of "accurate question answering - data extraction - visualization presentation". The technology lacks an effective natural language question answering interface, and users cannot directly obtain in-depth interpretation of the visualized data through question answering, resulting in extremely poor interaction flexibility.

[0015] (3) Lack of professional analytical capabilities. The relevant technologies are mostly limited to the basic visualization of data. They do not provide customized analytical functions in combination with the analytical logic of professional fields, and cannot provide professional analytical results with decision support value for industry users. There is a significant gap between them and the actual business needs of various industries.

[0016] (4) The system lacks self-evolution and dynamic routing capabilities. Each query requires reasoning from scratch using a large model, which makes it impossible to accumulate historical correct experience, resulting in serious waste of computing resources. Furthermore, the system lacks a hierarchical processing mechanism for queries of different complexities, leading to low system response efficiency.

[0017] In view of this, this application proposes an intelligent question-answering method based on large models and knowledge retrieval. By constructing a business database, it overcomes the knowledge blind spots of general large models in vertical business domains, addressing the problem of insufficient understanding of professional terminology and business logic in single-type knowledge bases, and achieving precise adaptation of professional knowledge to business scenarios. By constructing a relational business database and connecting it with business systems, it solves the problem of disconnect between professional knowledge and business databases in related technologies, and the large model's lack of understanding of database table structures and business entity relationships, resulting in low data query accuracy and inability to support complex business queries. By dynamically orchestrating modular query computing and visualization components, it solves the problem of poor synergy between "intelligent question answering" and "map spatiotemporal visualization," which prevents the realization of a seamless "question-answer intent recognition - accurate data query - visualization generation - result integration and output" chain. It addresses the problem of non-standardized visualization output in existing technologies by adhering to predefined standardized output specifications, ensuring the correctness of output content and normal interactive functionality through added validation. By introducing intent understanding, multi-level dynamic routing distribution, and an experience-based question-answering library, it achieves precise scheduling of computing resources and self-evolution of system capabilities, solving the problems of slow system response, high computing power consumption, and the inability to become smarter with use.

[0018] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0019] This application provides an intelligent question answering method based on large models and knowledge retrieval, referencing... Figure 1 ,include: Step 101: Obtain the user's input query statement and determine the user's intent based on the query statement.

[0020] Specifically, the process involves acquiring the user's input query, which can be in natural language. If the user's input is too concise, semantic expansion and entity extraction can be performed on the query. The query is then decomposed to uncover the user's true intent, providing foundational data for subsequent data matching and querying. When determining user intent, it's also necessary to determine if the user has a visualization intent—that is, whether they intend to visualize the response to the query. If so, the response information will need to be visualized subsequently.

[0021] Step 102: Search and match in a preset experience question and answer database according to the query statement, and determine the matching result and the matching degree of the matching result.

[0022] Specifically, the experience-based question and answer (Q&A) database is pre-built and stores multiple Q&A samples. Each sample includes historical query statements, corresponding correct and valid historical query statements, and historical visualization rendering data. This historical visualization rendering data provides users with intuitive interactive visualizations, allowing them to understand the spatiotemporal information of the visualized data. Subsequent queries can refer to detailed Q&A samples in the database, improving the accuracy and speed of system queries. As the system runs more frequently, the Q&A database grows, and the accuracy of the system's Q&A responses improves accordingly. The system records the "hit frequency" and "acceptance success rate" of each Q&A sample in the database in real time. For expired Q&A samples that have not been hit for a long time or have become invalid due to changes in the underlying business relational database table structure (Schema), the system performs periodic demotion or automated cleanup.

[0023] Therefore, each time a query is performed based on the query statement, multi-level dynamic routing is distributed according to the user's intent. Multi-level dynamic routing refers to dynamically determining the task's distribution path and the underlying resources invoked, and it is divided into two types of routing: long short-term memory routing and multi-level knowledge routing. Long short-term memory routing refers to performing high-dimensional vector matching on the user's input query statement in an experience-based question-and-answer database. First, the query statement is pre-matched in the experience-based question-and-answer database. The system uses the query statement input as a search term and performs a high-dimensional vector space search in the experience-based question-and-answer database to determine the matching results (i.e., historical query statements and historical visualization rendering data) and the matching degree. The matching degree refers to the similarity between the query statement and historical query statements in the experience-based question-and-answer database.

[0024] If the matching degree is greater than or equal to the first preset threshold (e.g., 95%), it is judged as a highly repetitive task, and short-circuit routing is performed. At this time, there is no need for subsequent large-scale model matching retrieval and related database query calculations. The matching result is directly returned as the query statement and visualization rendering data corresponding to the query statement. Executing the query statement yields the response information, improving the question-and-answer speed. Then, the response information and visualization rendering data are merged to obtain the response text. The response text is output so that the user can obtain the response content in natural language form. At the same time, interactive visualization data is presented to the user on the front end based on the visualization rendering data, improving the spatiotemporal visualization capabilities of business data with spatial attributes.

[0025] Step 103: In response to the matching degree being less than a first preset threshold, retrieve the context information corresponding to the query statement from the pre-built business database based on the query statement and the user intent.

[0026] Specifically, if the matching degree is less than the first preset threshold, it indicates that the current user-input query is not a highly repetitive task, requiring the triggering of a complex knowledge retrieval and large-scale model inference process via long-path routing. This triggers entry into multi-layered knowledge routing. If the matching degree is less than the first preset threshold (e.g., 95%) but greater than the second preset threshold (e.g., 60%), it indicates that the current matching result has medium-to-high confidence and some reference value. Therefore, the current matching result is used as auxiliary input information for the subsequent large-scale model, allowing it to directly mimic historical correct inference paths and parameter matching experience, bypassing complex inference steps from scratch and significantly improving the accuracy of generating correct query statements. By using the current matching result as auxiliary input information for the large-scale model, it directly constrains the model to use correct table names, field names, and relationships between tables, preventing the illusion of "fabricated tables / fields." Simultaneously, it constrains the model to use correct business indicator calculation rules, reusing the calculation paradigm of the example and eliminating computational logic illusions. It also directly constrains the syntax format and query structure of subsequently generated query statements, avoiding problems such as syntax errors, missing fields, and structural confusion. The current matching results also allow the large model to follow business intent and indicate the correct reasoning path for data queries. Subsequently, the large model will imitate the reasoning path in the matching results, without deviating from the business objectives, and without querying irrelevant tables or irrelevant data.

[0027] However, if the matching degree is less than the second preset threshold, it means that the current matching result has low confidence and is not of reference value, and therefore should not be used as auxiliary input information for subsequent large models.

[0028] Multi-layered knowledge routing is performed, which means retrieving contextual information from a pre-built business database based on the query and user intent. This contextual information serves as input prompts for the large-scale model, enabling it to determine the accurate response.

[0029] The business database comprises a vector database and a graph database. These databases primarily provide background knowledge based on user questions, reducing blind spots in the knowledge base of the large model. The vector database stores all text query content within the business domain, allowing users to retrieve conceptual information, professional knowledge, and other relevant content. For example, the satellite communication field contains many technical terms such as satellite, beam, and transponder. These terms are not readily understood by the general large model, thus requiring the vector database for querying. Furthermore, to facilitate the large model's understanding of specific database table structures, the vector database supplements it with the pre-built relational database table content and explanations, as well as the relationships between tables. This helps the large model understand which database tables to use. When calculating data metrics, if relationships between multiple tables are involved, the large model can determine how to perform these relationships and metric calculations.

[0030] Graph databases consist of triples, containing business entity objects, their attributes, and the relationships between them. This helps large models better understand business data, overcoming the limitations of answering queries solely based on textual information. For example, in satellite communications, taking a communication satellite as an example, the satellite is the platform for communication missions, and the payload is its core functional module, containing transponders and antennas. The transponder receives ground signals, amplifies them, and then forwards them; the antenna forms a directional satellite beam to achieve precise signal coverage. Satellites, payloads, transponders, antennas, and beams are considered business entity objects. The specific relationships between each business entity object, as well as the types of these relationships, can be constructed using a graph database. The graph database also stores the physical table names, core field names, and primary / foreign key relationships corresponding to the business entity objects in the relational database. When parsing user intent, the correct physical table structure and the shortest path can be obtained by following the graph path in the graph database, reducing model illusion (model illusion refers to the model fabricating table names or using incorrect relationships). In addition, the graph database also stores the calculation rules for business metrics, which are written into the graph database as special attributes or rule nodes of business entity objects. When a user's query involves this metric, the large model does not need to deduce the mathematical logic itself; it directly extracts the calculation rules from the graph database and generates the query, eliminating the "illusion of computational logic" of the large model.

[0031] Relational databases interface with business systems to enable scheduled updates of static data and real-time updates of dynamic data for real-world business operations. Static data refers to records of core business data, while dynamic data is the latest status information of real-time data. For example, in the field of satellite communication, static data includes satellite tables and beam tables. Satellite tables record basic information about each satellite, and beam tables record information such as the coverage area of ​​each beam. Dynamic data includes status data that changes over time, such as the number of users currently using a particular beam and the specific uplink and downlink bandwidth used by each user for that beam. To facilitate subsequent database table retrieval and data querying based on large models, each database table needs to be described and annotated in Chinese during the relational database construction phase. For each database, the business themes contained within it and its intended use must be accurately described. For each data table, the core business objects, update frequency, and key role played in the business process must be accurately described. For each field in a data table, the actual business meaning of the field must be accurately described through comments. For numerical values, the range of values ​​must be described. The status information field needs to describe all existing status information.

[0032] Furthermore, if the user's query is a concept definition task (such as "What is a repeater?"), it is routed directly to the vector database, outputting lightweight text as the response. If the user's query is not a concept definition task, multi-layered knowledge routing includes routing to the vector database, graph database, and relational database. The task is distributed through a tiered retrieval route using "vector anchoring" in the vector database, "graph navigation" in the graph database, and "data retrieval from tables" in the relational database.

[0033] Furthermore, the step of retrieving the context information corresponding to the query statement from a pre-built business database based on the query statement and the user intent includes: Based on the query statement and the user intent, key business information and business database table structure information are retrieved from the vector database; based on the query statement and the user intent, business entity information and attribute information are retrieved from the graph database; the key business information, the business database table structure information, the business entity information, and the attribute information are used as the context information.

[0034] Specifically, based on the query and user intent, the system retrieves and matches vertical industry-specific business knowledge from the vector database to identify key business information. Simultaneously, the vector database stores database table structure knowledge, including information on all database tables (database name, table name, field meaning explanations, and inter-table field relationships). This information is then retrieved and matched to determine the business database table structure. Furthermore, based on the query and user intent, the system queries the graph database to obtain business entity and attribute information. This key business information, business database table structure information, business entity information, and attribute information are used as contextual information and subsequently input into the larger model. This allows the larger model to output accurate query statements, improving the accuracy of subsequent responses. By constructing a three-layer knowledge system of vector database, graph database, and relational database, the system effectively solves the knowledge blind spots and "illusion" problems inherent in general-purpose large models within specialized fields. The large model can dynamically understand business terminology, logical relationships, and underlying data structures, thereby generating accurate and executable complex business queries (such as multi-table joins and indicator calculations), achieving a leap from "shallow dialogue" to "deep data exploration."

[0035] Step 104: Based on the user intent, the context information, and / or the matching result, determine at least one corresponding task flow component and the execution order of each task flow component using the large model.

[0036] Furthermore, based on the user intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model, including: In response to the matching degree being less than a first preset threshold and greater than a second preset threshold, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model based on the user intent, the context information, and the matching result; in response to the matching degree being less than the second preset threshold, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model based on the user intent and the context information.

[0037] Specifically, if the matching degree is less than a first preset threshold but greater than a second preset threshold, indicating that the matching result has some reference value, the user intent, the context information, and the matching result are input into the large model. The large model then orchestrates at least one task flow component and the execution order of each task flow component. The task flow components include query calculation components and visualization components. At least one task flow component includes at least one query calculation component and one visualization component. The at least one query calculation component can have a serial query relationship or a parallel query relationship. The large model determines whether there are data dependencies between different query calculation components based on the business entity information and attribute information in the context information. If data dependencies exist, the dependency path is determined, and thus the execution order of different query calculation components is determined. If no data dependencies exist, multiple query calculation components can be queried in parallel to reduce latency. Furthermore, if the execution of the visualization component depends on the output of the query calculation component, the query calculation component is executed first, and then the output of the query calculation component is used as the input parameter for the visualization component, which is then executed.

[0038] The matching results input into the large model can help it determine the correct task flow components and the execution order of each component. These matching results, as few-shot examples, are injected into the large model, allowing it to directly mimic historically correct inference paths, database and table call relationships, parameter configuration rules, and output format specifications during subsequent query generation and visualization configuration. This enables precise constraints on the large model's output behavior.

[0039] If the matching degree is less than the second preset threshold, it indicates that the matching result is not reliable. In this case, the user intent and the context information are input into the large model, and at least one task flow component and the execution order of each task flow component are obtained through orchestration by the large model. The orchestration method is the same as when the matching result is reliable, and will not be repeated here.

[0040] Through the dynamic orchestration of task flow components in this embodiment, data dependencies in complex business processes can be precisely adapted, ensuring the correctness of query logic. Parallel execution of dependency-free task flow components is supported, significantly improving system response speed. Decomposing the overall complex task into multiple modular task flow components reduces the complexity of large model inference and minimizes illusions and errors. When executing different task queries, task flow components can be added, deleted, replaced, and reorganized as needed without modifying the core architecture. This allows for the rapid construction of intelligent question-answering and spatiotemporal visualization processes for corresponding domains, demonstrating extremely strong cross-industry adaptability.

[0041] Step 105: Execute each task process component according to the execution order to obtain response information and visualization rendering data.

[0042] Furthermore, the step of executing each task process component according to the execution order to obtain response information and visualization rendering data includes: executing each query calculation component according to the execution order to obtain the response information; and executing the visualization component based on the response information to obtain the visualization rendering data.

[0043] Specifically, when it is determined that the visualization component needs to rely on the output data of the query calculation component, the query calculation component is executed first, followed by the visualization component.

[0044] The process of executing each query calculation component in the order of execution to obtain the response information includes: generating a corresponding query statement for each query calculation component through a large model, and executing the query statement based on a pre-built relational database to obtain the response information.

[0045] A single query computation component can only query one database. If multiple databases need to be queried, multiple query computation components need to be orchestrated, and the order of these components must be determined. If there is no order of query computation components determined by the user's intent, parallel queries can be used, meaning multiple query computation components execute in parallel to improve query speed. For example, when querying all playgrounds and hospitals in a map area, two business objects (playgrounds and hospitals) are represented in different tables in two databases. Two query computation components are generated according to the user's query intent and specified query conditions, and database connections and query operations are performed. Because there is no order of query components, the two query computation components can perform parallel queries.

[0046] However, if there is a sequential relationship between the query content, the queries must be performed according to this relationship, and data transfer is required between adjacent query calculation components. That is, the output data of the first executed query calculation component becomes the input data for the later executed query calculation component. For example, in a map-based satellite communication service scenario, it is necessary to query the effective coverage population of the B1 area under a certain satellite S1 beam. In this case, the data queries have a sequential relationship. If the satellite beam coverage area and the population statistics are in two separate databases, then a query calculation component AC1 for querying the effective coverage area of ​​the B1 area of ​​the satellite S1 beam must first be programmed. Based on the query results of AC1 (the specific latitude and longitude spatial range), another query calculation component AC2 needs to be programmed to query the resident population of this area based on the query results of AC1, ultimately obtaining the result.

[0047] For each query calculation component, when executing the component, its task information and context information are input into the large model. The large model then generates a query statement, which in this embodiment is SQL (Structured Query Language). The query statement is executed to retrieve data from the relational database and calculate corresponding metrics, yielding a response. It should be noted that the data content and format of the response information obtained after the task flow component executes the query statement must be defined in advance. The defined data content and format of the response information include: what content data the response information should contain, and the output format of the content data, such as an output table where the first row contains field attribute information, etc.

[0048] The query statements generated by the large model include both single-database, single-table queries and cross-table queries across multiple databases. When executing a query, the name of the database to be invoked must be explicitly specified, and the specified database query tool is invoked based on the database name to execute the query.

[0049] Furthermore, the step of executing the visualization component based on the response information to obtain the visualization rendering data includes: Based on the response information and the task information of the visualization component, the business object corresponding to the visualization component is extracted through a large model; based on the business object and the response information, the visualization rendering data is generated.

[0050] Specifically, the visualization component is used to realize visualization-related business requirements and content. From the user's query and the retrieved data, it extracts business objects that can be represented in a spatiotemporal form and performs interactive spatiotemporal map visualization. Through the structured integration and format conversion of spatial positioning information and static and dynamic business objects, it generates visualization rendering data that conforms to the front-end's agreed standards, so as to support the front-end to directly parse the visualization rendering data and complete the visualization rendering and interactive display of the map.

[0051] In practice, based on the response information and the task information of the visualization components, the large model extracts the business objects corresponding to the visualization components. The response information contains real business data corresponding to the user's query, covering the basic attributes, numerical information, and spatial association information of the business objects. For example, if a user queries "the distribution of central stations and the number of online users within the B1 area of ​​a certain satellite's S1 beam," the response information will include the latitude and longitude range of the beam's coverage area, the ID, name, specific location, and equipment status of all central stations within the area, as well as the number of online users corresponding to each central station. This data is the core basis for identifying business objects. The task information of the visualization components clarifies the component's working objectives, extracting business objects that can be represented in a spatiotemporal form, while defining the screening scope, type requirements, and association rules for business objects. The large model first performs semantic parsing on the response information, extracting all the business entities contained therein, and then combines it with the task information of the visualization components to filter out business objects that meet the "spatiotemporally representable" condition—that is, entities that have spatial attributes (such as latitude and longitude, spatial coverage area) and can be displayed through map elements (points, lines, areas, trajectories). Simultaneously, the large model categorizes and deduplicates the extracted business objects, clarifying the core attributes of each object and providing a foundation for subsequent visualization rendering data generation. For example, targeting a specific application area of ​​satellite communication, combining user queries, response information, and task information from visualization components, the large model can accurately extract the following visualization business objects: (1) Fixed spatial entities: such as central stations, remote stations, satellite ground receiving terminals, etc. These business objects have fixed latitude and longitude coordinates and can be displayed on the map as points. Their core attributes include device ID, name, location, working status, bandwidth usage, etc. (2) Spatial coverage area: such as satellite beam coverage area, central station signal coverage area, etc. These types of business objects have a clear spatial range and can be displayed on the map in the form of a surface. Their core attributes include the latitude and longitude range of the coverage area, coverage area, signal strength, etc. (3) Dynamic business objects: such as mobile terminals, satellite trajectories, etc. These business objects have dynamically changing spatial locations and can be displayed on the map in the form of trajectories. Their core attributes include object ID, latitude and longitude sequence of the movement trajectory, movement speed, real-time status, etc.

[0052] It should be noted that during the extraction of business objects, the large model will strictly follow the task information of the visualization components to avoid extracting business objects that are irrelevant to the visualization (such as purely numerical statistical results, business process descriptions, etc.). At the same time, it will ensure that the extracted business objects are completely consistent with the response information, without adding irrelevant objects or omitting core objects, thus ensuring the accuracy of the subsequent visualization rendering data.

[0053] Based on the extracted business objects and response information, generating visualization rendering data is the core output of the visualization component. Its core is to structurally integrate and format-convert the extracted business objects and their associated data, generating standardized data that conforms to the front-end's agreed standards. This ensures that the front-end map rendering engine can directly parse and render the data, while simultaneously achieving deep integration of business data and map interaction. The large model associates and binds each extracted business object with the corresponding core data in the response information, clarifying the attribute information, spatial information, and relationships of each business object. For example, the business object of a central station is bound to data such as its latitude and longitude, equipment status, and number of online users in the response information; the business object of a satellite beam coverage area is bound to data such as the latitude and longitude range, signal strength, and number of covered users in the response information, ensuring the completeness and accuracy of the information for each business object. For each business object with spatial attributes, the large model extracts its spatial positioning information (latitude and longitude, spatial range, trajectory sequence, etc.) and performs structured processing according to the format agreed upon by the front end.

[0054] Visualization rendering data must meet standardized output specifications. These specifications are the agreed-upon visualization rendering data format standards with the front-end, including field definitions for map perspectives (naming and format requirements for longitude, latitude, and altitude perspective fields), list structure standards for static and dynamic elements (mandatory and format constraints for key fields such as element identifier, name, and type), GeoJSON data format standards, and configuration specifications for rendering surface styles and interactive pop-ups. The output format requirements define the core content and standardized format of the map visualization rendering configuration data output by this component. Visualization rendering data output according to the visualization data format standards is a structured JSON (JavaScript Object Notation) object containing three core modules: ① Basic positioning parameter module (including spatial positioning fields such as center point latitude and longitude, and altitude perspective); ② Static element module (containing a list of elements conforming to the agreed-upon structure, each element containing basic attributes, rendering surface configuration, and pop-up attribute information; rendering surface data supports GeoJSON or latitude / longitude list format); ③ Dynamic element module (containing a list of elements conforming to the agreed-upon structure, each element containing basic attributes, trajectory route, movement speed, and pop-up attribute information). The front-end can directly parse this JSON object to complete map rendering and interaction. The structured JSON object can be directly recognized, parsed, and rendered by various map rendering engines without additional format conversion, greatly improving the development efficiency and adaptability of visualization.

[0055] The basic positioning parameter module is the core positioning unit for visualization. Its function is to provide users with rapid spatial area positioning, clarifying the display range and viewing parameters of the entire spatiotemporal map information. Its data structure is centered on the center point and includes key positioning parameters such as longitude (lng), latitude (lat), page height / viewpoint (alt), heading angle, and pitch angle. Among these, latitude and longitude parameters are used to accurately determine the geographic coordinates of the spatial area. Height / viewpoint, heading angle, and pitch angle together are responsible for the overall screen positioning, accurately defining the display range and viewing angle of spatiotemporal information. This ensures that users can quickly locate the target spatial area, providing a unified and accurate spatial reference for the subsequent visualization rendering of static and dynamic elements.

[0056] The static element module is used to render fixed geographic objects in the spatiotemporal map (such as poles and cameras in sensing nodes). It organizes data using a standardized list format. The front-end can quickly identify object categories through element type and determine the icon resources displayed on the map through the icon path (imgURL, image Uniform Resource Locator), achieving accurate rendering and categorized display of static elements. Each data entry includes id, name, type, latitude and longitude, icon, renderSurface, and attributes. The specific design is as follows: ① Basic Information Fields: Element ID is a unique identifier for static elements, ensuring that each element can be identified and manipulated individually; Name field is used to label the element's identity, making it easy for users to distinguish them intuitively; Type field (such as camera, rod, sensor) clarifies the element category, providing a basis for judgment in the front-end rendering logic; Latitude and Longitude field (location) determines the specific geographical coordinates of the element on the map; Icon Path (imgURL) specifies the icon resources displayed by the element, ensuring the consistency and standardization of the visualization presentation.

[0057] ② RenderSurface: The render surface is an extended visual area of ​​a static element, used to display the element's associated spatial information (such as camera coverage area, area affected by poles, etc.). Its data structure includes three core parts: type, geometric data, and style. The geometric data supports GeoJSON format or latitude / longitude lists, adapting to various geometric shapes such as points, lines, and polygons; the type field specifies the concrete shape of the render surface (such as circle, polygon, etc.); the style field configures visual parameters such as the render surface's color scheme and border width, ensuring clear and discernible visualization.

[0058] ③ Attributes: The attribute field is used to store the unique parameters of static elements (such as the day and night coverage distance of the camera, device status, etc.). The front end can display the core attributes contained in this field through a mouse hover pop-up window, so that users can quickly obtain detailed information about the element and improve the information interaction capability of the spatiotemporal map.

[0059] The dynamic element module is used to render geographic objects with movement characteristics in a spatiotemporal map. Its basic feature attributes are consistent with those of static elements to ensure data format uniformity and compatibility. At the same time, a new trajectory route-related field is added to realize the visualization of the movement of dynamic elements. The specific design is as follows: ① Basic information fields: Consistent with static elements, including core fields such as id, name, type, icon path (imgURL), latitude and longitude (location). The name field will be labeled next to the chart to help users quickly identify the specific object corresponding to the dynamic element; the latitude and longitude field (location) is used to determine the initial position of the dynamic element.

[0060] ② Track and route related fields: The track and route are defined by the route field, which contains a series of continuous latitude and longitude coordinates. The dynamic element will move according to this coordinate sequence. The speed field is used to control the movement rate, which is in milliseconds. It defines the number of milliseconds at which the dynamic element moves forward one path point, so as to realize precise control and visualization of the movement process of the dynamic element.

[0061] ③ Attributes: Consistent with the design of attribute fields for static elements, these fields store unique parameters of dynamic elements (such as the earliest appearance time of the object, identification intent, feature description, etc.). They support viewing via mouse hover pop-ups, providing users with detailed background information on dynamic elements and assisting them in interpreting and analyzing dynamic spatiotemporal events.

[0062] The method described in this embodiment ensures that the visualized content closely matches user needs and business data. Business object extraction is based on response information (real business data) and component task information (visualization requirements), avoiding the extraction of irrelevant objects and ensuring that the visualized content accurately matches the user's inquiry intent. Standardized output specifications generate rendering data that conforms to front-end agreements, eliminating the need for additional front-end processing and improving visualization rendering efficiency and adaptability. By classifying static and dynamic business objects, rendering styles and interaction rules are clearly defined, supporting the front-end to implement various interactive functions such as point-to-point clicking, floating pop-ups, and trajectory playback, enhancing the user experience. The standardized rendering data format is adaptable to different types of map rendering engines, eliminating the need to modify the data format for different engines and improving the system's scalability and compatibility.

[0063] Step 106: Merge the response information and the visualization rendering data to obtain the response text corresponding to the query statement.

[0064] Specifically, based on the response information and visualization rendering data, a specific prompt word template (such as spatiotemporal interactive prompt words, requiring the output of a large model in a specified format) is used. Data fusion is performed through the large model to generate response text as a natural language response to the user.

[0065] Step 107: Output the response text and perform a visual interactive display based on the visualization rendering data.

[0066] Specifically, the output response text serves as the final answer. During the interactive visualization, a GIS (Geographic Information System) map is used, and the visualization rendering data (structured JSON objects) is displayed interactively on the front-end page according to a specified format. This achieves alignment and integration of the response text and the visualization rendering data, ensuring that the final output includes natural language conclusions and standardized map rendering instructions, realizing a collaborative output of "one question, two answers" (text + map).

[0067] Based on steps 101 to 107 above, the intelligent question-answering method based on a large model and knowledge retrieval provided in this application includes: acquiring a user-inputted query statement, and determining the user's intent based on the query statement. A search and matching process is performed in a preset experience question-answering database based on the query statement to determine the matching result and the matching degree of the matching result. The experience question-answering database can provide historical and accurate experience to increase the accuracy of the large model's response. As the system's operating frequency increases, the accumulated content in the experience question-answering database gradually increases, and the accuracy of the system's response also gradually increases. If the matching degree is less than a first preset threshold, indicating that no matching result with a high matching degree is found in the experience question-answering database, then the context information corresponding to the query statement is retrieved from a pre-constructed business database based on the query statement and the user's intent. The business database stores domain knowledge closely related to the business domain, which can be used as context information for the subsequent large model to provide precisely adapted content. Based on the user's intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined through the large model, that is, the planning of the automated link is realized through the large model. By analyzing the business exploration logic, the execution order of each task process component can be determined, which helps to accurately determine the corresponding response information for the query statement. The modular design of the task process components and the clear context information enable the system to quickly adapt to different vertical industries through configuration and knowledge updates. By flexibly arranging the execution order of the task process components, it can handle various business scenarios ranging from simple queries to complex multi-step analyses. Each task process component is executed according to the execution order to obtain response information and visualization rendering data. The response information and the visualization rendering data are fused to obtain the response text corresponding to the query statement. The response text is output and displayed interactively based on the visualization rendering data. This application not only provides users with response text matching the query statement, but also provides users with more intuitive interactive visualization information, improving the spatiotemporal visualization capabilities of business data with spatial attributes.

[0068] In some embodiments, the method further includes: The large model is used to verify the response information and the visualization rendering data respectively. In response to the response information failing the verification or the visualization rendering data failing the verification, the query statement is regenerated based on the verification failure information.

[0069] Specifically, to ensure the accuracy of the final response text and visualization, both the response information and the visualization rendering data need to be validated separately. When validating the response information, the large model validates the response information based on the task information of the query calculation component to determine if the query statement was executed correctly. Response information failing validation includes either containing error messages or, although a response is returned, it does not meet the validation conditions. Failure to meet validation conditions includes responses that do not match the task information of the query calculation component, or responses that are not output according to the pre-defined data content and format. If the large model determines that the response information has failed validation, it needs to provide a validation failure message including the reason for the failure. Adding validation ensures the correctness of the response content, guarantees that the large model can correctly output content in the user-specified format, and improves the standardization of the response information.

[0070] If the validation fails, the query statement is regenerated. In practice, based on the task information and context information of the query calculation component, as well as the validation failure information generated from the large model, a new query statement is generated using the large model. Then, the query calculation component is invoked to execute the regenerated query statement, obtaining new response information, until the new response information passes validation. Otherwise, new query statements are continuously generated. If the validation passes, the execution process of the query calculation component ends.

[0071] Similarly, when validating visualization rendering data, the large model validates the data based on the task information of the visualization component to determine its correctness. Failure to validate includes receiving error messages, or, although data is returned, it not meeting the validation criteria. Failure to meet validation criteria includes mismatches between the visualization rendering data and the task information of the visualization component, or outputting visualization rendering data that does not conform to the pre-defined data content and format. If the large model determines that the visualization rendering data has failed validation, it needs to provide validation failure information, including the reason for the failure. Adding validation ensures the correctness of the visualization rendering data content, guarantees the large model can correctly output content in the user-specified format, and improves the standardization of the visualization rendering data.

[0072] Furthermore, to ensure the correctness of the interactive visualization, it is necessary to determine whether the response information includes all the generated elements required by the visualization component. The business object corresponding to the visualization component is extracted using a large model. The extraction method for the business object is the same as in the previous embodiment and will not be repeated here. Based on the business object and context information, the large model determines whether the response information includes the generated elements of the business object. For example, generated elements include the latitude and longitude coordinates of the business object, key attribute information of the business object, etc. If the business object has surrounding rendering space data, it is also necessary to determine whether the generated elements include the spatial location information of the surrounding rendering space. If it is determined that the response information does not include all the generated elements required by the business object, a secondary query is required. Specifically, based on the task information and context information of the query calculation component, as well as the information on the generated elements required by the business object, the query statement is regenerated using the large model. The query calculation component is called to execute the regenerated query statement to obtain the regenerated response information, ensuring that the regenerated response information includes the generated elements required by the business object. Through the secondary query process in this embodiment, it can be ensured that the response information includes all the generated elements required by the business object, thereby ensuring the correctness of the subsequent visualization.

[0073] This application starts with user intent recognition and, through intelligent task flow orchestration, connects multiple stages such as business database retrieval, task flow determination, data querying, and map generation into an organic whole. The system can automatically determine whether visualization is required and dynamically orchestrate task flow components with / without data dependencies, ultimately outputting semantically consistent response text and spatiotemporal visualization results, achieving end-to-end closed-loop interaction. By introducing a verification mechanism (verification of response information and visualization rendering data) and a secondary query mechanism, the system ensures accurate mapping from raw data to visualization elements. Standardized output specifications allow the generated visualization rendering data to be directly parsed and rendered by the front end, supporting rich front-end visualization interactions (such as floating pop-ups), solving the problems of unreliable visualization results and poor interactivity in existing technologies.

[0074] Meanwhile, this application employs a modular, component-based design (such as query calculation components and visualization components) and clearly defined contextual information, enabling the system to quickly adapt to different vertical industries (such as satellite communication, smart cities, and logistics scheduling) through configuration and knowledge base updates, without modifying the core architecture. The flexibility of task flow orchestration also allows it to handle various business scenarios, from simple queries to complex multi-step analyses. The system directly connects to the real-time and historical databases of business systems, ensuring that all question-and-answer and visualization content is based on real and reliable data sources. Combined with the association analysis capabilities of the domain knowledge graph in the graph database, it can provide users with deeper insights and decision-making basis that go beyond surface data and contain business logic, truly achieving an upgrade from "data display" to "business analysis."

[0075] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0076] It should be noted that some embodiments of this application have been described above. In some cases, the actions or steps described in the above embodiments can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0077] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a question-answering device based on a large model.

[0078] refer to Figure 2 The intelligent question-answering device based on large models and knowledge retrieval includes: The acquisition module 201 is configured to acquire a query statement input by the user and determine the user's intent based on the query statement; The matching module 202 is configured to perform a search and matching in a preset experience question and answer database based on the query statement, and determine the matching result and the matching degree of the matching result; The retrieval module 203 is configured to, in response to the matching degree being less than a first preset threshold, retrieve the context information corresponding to the query statement from a pre-built business database based on the query statement and the user intent; The determination module 204 is configured to determine at least one corresponding task flow component and the execution order of each task flow component based on the user intent, the context information and / or the matching result through a large model. Execution module 205 is configured to execute each task process component in the execution order to obtain response information and visualization rendering data; The fusion module 206 is configured to fuse the response information and the visualization rendering data to obtain the response text corresponding to the query statement; The visualization module 207 is configured to output the response text and perform a visual interactive display based on the visualization rendering data.

[0079] In some embodiments, the business database includes a vector database and a graph database; the retrieval module 203 is configured to query the vector database to obtain key business information and business database table structure information according to the query statement and the user intent; query the graph database to obtain business entity information and attribute information according to the query statement and the user intent; and use the key business information, the business database table structure information, the business entity information, and the attribute information as the context information.

[0080] In some embodiments, the determining module 204 is configured to, in response to the matching degree being less than a first preset threshold and greater than a second preset threshold, determine at least one corresponding task flow component and the execution order of each task flow component through a large model based on the user intent, the context information, and the matching result; and in response to the matching degree being less than the second preset threshold, determine at least one corresponding task flow component and the execution order of each task flow component through a large model based on the user intent and the context information.

[0081] In some embodiments, the at least one task flow component includes at least one query calculation component and a visualization component; the execution module 205 is configured to execute each query calculation component in the execution order to obtain the response information; and to execute the visualization component based on the response information to obtain the visualization rendering data.

[0082] In some embodiments, the execution module 205 is configured to generate a corresponding query statement for each query calculation component using a large model, and execute the query statement based on a pre-built relational database to obtain the response information.

[0083] In some embodiments, the execution module 205 is configured to extract the business object corresponding to the visualization component through a large model based on the response information and the task information of the visualization component; and generate the visualization rendering data based on the business object and the response information.

[0084] In some embodiments, the apparatus further includes a verification module configured to verify the response information and the visualization rendering data using a large model, and to regenerate the query statement based on the verification failure information if the response information fails verification or the visualization rendering data fails verification.

[0085] In some embodiments, the retrieval module 203 is configured to, in response to the matching degree being greater than or equal to a first preset threshold, obtain the response information and visualization rendering data corresponding to the query statement based on the matching result.

[0086] In some embodiments, an update module is also included, configured to update the preset experience question and answer database based on the response information, the visualization rendering data, and the query statement.

[0087] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0088] The apparatus described above is used to implement the corresponding intelligent question answering method based on large models and knowledge retrieval in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0089] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the intelligent question-answering method based on large model and knowledge retrieval as described in any of the above embodiments.

[0090] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0091] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0092] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0093] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0094] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0095] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0096] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0097] The electronic devices described above are used to implement the corresponding intelligent question-answering methods based on large models and knowledge retrieval in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0098] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the intelligent question-answering method based on large models and knowledge retrieval as described in any of the above embodiments.

[0099] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0100] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the intelligent question answering method based on large model and knowledge retrieval as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0101] Based on the same concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to perform the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0102] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0103] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to choose, based on the prompt message, whether to provide personal information to the software or hardware such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution.

[0104] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" regarding the provision of personal information by the electronic device.

[0105] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0106] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0107] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0108] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0109] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. An intelligent question-answering method based on large models and knowledge retrieval, characterized in that, include: Obtain the user's input query statement and determine the user's intent based on the query statement; The search and matching are performed in a preset experience question and answer database based on the query statement to determine the matching results and the matching degree of the matching results; In response to the matching degree being less than a first preset threshold, the context information corresponding to the query statement is retrieved from a pre-built business database based on the query statement and the user intent; Based on the user intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model; Each task process component is executed in the described execution order to obtain response information and visual rendering data; The response information and the visualization rendering data are fused to obtain the response text corresponding to the query statement; The response text is output and a visual interactive display is performed based on the visualization rendering data.

2. The method according to claim 1, characterized in that, The business database includes a vector database and a graph database; The step of retrieving the context information corresponding to the query statement from a pre-built business database based on the query statement and the user intent includes: Based on the query statement and the user intent, key business information and business database table structure information are obtained by querying the vector database. Based on the query statement and the user intent, business entity information and attribute information are obtained by querying the graph database; The key business information, the business database structure information, the business entity information, and the attribute information are used as the context information.

3. The method according to claim 1, characterized in that, Based on the user intent, the context information, and / or the matching result, at least one corresponding task flow component and the execution order of each task flow component are determined using a large model, including: In response to the matching degree being less than a first preset threshold and greater than a second preset threshold, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model based on the user intent, the context information, and the matching result. In response to the matching degree being less than the second preset threshold, at least one corresponding task flow component and the execution order of each task flow component are determined through a large model based on the user intent and the context information.

4. The method according to claim 1, characterized in that, The at least one task flow component includes at least one query calculation component and a visualization component; the execution of each task flow component according to the execution order to obtain response information and visualization rendering data includes: Each query calculation component is executed in the execution order described above to obtain the response information; The visualization component is executed based on the response information to obtain the visualization rendering data.

5. The method according to claim 4, characterized in that, The step of executing each query calculation component according to the execution order to obtain the response information includes: For each query calculation component, a corresponding query statement is generated through the large model, and the query statement is executed based on a pre-built relational database to obtain the response information.

6. The method according to claim 4, characterized in that, The step of executing the visualization component based on the response information to obtain the visualization rendering data includes: Based on the response information and the task information of the visualization component, the business object corresponding to the visualization component is extracted through the large model; The visualization rendering data is generated based on the business object and the response information.

7. The method according to claim 5, characterized in that, The method further includes: The large model is used to verify the response information and the visualization rendering data respectively. In response to the response information failing the verification or the visualization rendering data failing the verification, the query statement is regenerated based on the verification failure information.

8. The method according to claim 1, characterized in that, The method further includes: In response to the matching degree being greater than or equal to a first preset threshold, the response information and visualization rendering data corresponding to the query statement are obtained based on the matching result.

9. The method according to claim 1, characterized in that, The method further includes: The preset experience question and answer database is updated based on the response information, the visualization rendering data, and the question statement.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 9.

Images