CN-121996770-A - Method and system for relieving illusion of large model

Abstract

The invention relates to a method and a system for relieving illusion of a large model, and belongs to the technical field of computers. The method comprises the steps of judging whether a local knowledge base, a search engine or a large model self capability is called to obtain an answer through intention recognition analysis of a problem input by a first-level agent, if the search engine is called, pulling a preset number of web pages related to the content of the problem queried by the search engine through the first-level agent, inputting the content, web address information and problem and prompt words input by the user into the large model to obtain the answer, if the local knowledge base is called, analyzing the problem through a second-level agent and generating query instructions corresponding to the problem, and performing interactive query on the query instructions and professional databases of industries to which at least two problems with different query modes belong to obtain a query result, inputting the problem input by the user, the prompt words and the query result into the large model to obtain the answer, and if the large model self capability is called, obtaining the answer according to the problem through the large model.

Inventors

• XU DAN
• HUANG LIJUN
• GAO FANGYU
• LI XIANWEI
• Tong Lianhui
• LI YINGHAO
• ZHANG YANLONG
• LV YONGQIANG
• HE XINGDUO
• LU SHENG
• LI MINGDAO
• ZHANG XU

Assignees

• 许继电气股份有限公司
• 许昌许继软件技术有限公司
• 中国电气装备集团有限公司

Dates

Publication Date

20260508

Application Date

20241108

Claims (5)

1. 1. A method for relieving illusion of a large model is characterized by comprising the steps of judging whether a local knowledge base, a search engine or the capacity of the large model is called through intention recognition analysis of a first-level agent to a problem input by a user to obtain an answer, pulling a preset number of web pages related to content which is inquired by the problem through the search engine through the first-level agent if the search engine is called, inputting the content, web address information of each web page and the problem and a prompt word which are input by the user into the large model to obtain the answer, analyzing the problem through a second-level agent and generating an inquiry command corresponding to the problem if the local knowledge base is called, carrying out interactive inquiry on the inquiry command and a professional database of industries which at least two inquiry modes are different from each other to obtain an inquiry result, inputting the problem, the prompt word and the inquiry result which are input by the user into the large model to obtain the answer, and obtaining the answer according to the problem through the large model if the capacity of the large model is called.
2. 2. The method of claim 1, wherein the step of interactively querying at least two databases comprises querying any or all of the databases simultaneously by a query instruction and outputting corresponding query results.
3. 3. The method of claim 1, wherein the step of interactively querying at least two databases comprises sequentially interactively querying the query instructions in database order, wherein the database order is custom-organized.
4. 4. The method of claim 1, wherein the specialized databases of the industry to which the problem belongs include a knowledge graph database, an SQL relational database, and a RAG vector database for storing specialized data of the industry to which the problem belongs.
5. 5. A system for mitigating large model illusions comprising a processor, wherein the processor is configured to execute a computer program to implement the steps of the method for mitigating large model illusions of any of claims 1-4.

Description

Method and system for relieving illusion of large model Technical Field The invention belongs to the technical field of computers, and particularly relates to a method and a system for relieving illusion of a large model. Background Currently, the general large model is applied to a plurality of different fields and tasks, but the illusion of fact knowledge conflict often occurs facing a vertical scene, and the application of the industry is affected. The operation and maintenance of the electric power equipment comprise equipment inspection, overhaul maintenance, fault processing, performance testing and the like, the state monitoring and fault diagnosis need to ensure the accuracy and the specialty, and the generated large model often generates illusions due to lack of corresponding industry knowledge, so that errors are answered. The existing mode is to alleviate the illusion of the large model through the assistance of a vector knowledge base, and particularly, the large model is referred and answered through recall of vector similarity to knowledge segments related to questions. However, this approach has three problems (1) the user's problem may require cross-document summarization, while vector libraries are difficult to recall the relevant knowledge segments accurately and completely. (2) And (3) for some equipment information and parameter information, the vectorization model is difficult to map to semantic space accurately for the numerical information, so that recall failure is caused. Disclosure of Invention The invention aims to provide a method and a system for alleviating the illusion of a large model, which are used for solving the problems that in the prior art, only a vector library is used, so that the latest information is difficult to acquire in real time, recall incompleteness and recall failure are easy to occur. The invention provides a method for relieving illusion of a large model, which comprises the steps of judging whether a local knowledge base, a search engine or the capacity of the large model is called through intention recognition analysis of a first-level agent on a problem input by a user, if the search engine is called, pulling a webpage of which the content is related to the content which is queried by the problem through the search engine through the first-level agent, inputting the content, website information and the problem input by the user and a prompt word of the large model to obtain an answer, if the local knowledge base is called, analyzing the problem and generating a query instruction corresponding to the problem through a second-level agent, obtaining a query result through interactive query of the query instruction and a professional database of industries to which at least two problems with different query modes belong, inputting the problem input by the user, the prompt word and the query result into the large model to obtain the answer, and if the capacity of the large model is called, obtaining the answer according to the problem through the large model. Further, the method for interactive query with at least two databases comprises the steps of simultaneously querying any one or all databases by query instructions and outputting corresponding query results. Further, the method for interactive query with at least two databases comprises the steps of sequentially carrying out interactive query on query instructions according to the sequence of the databases, and carrying out self-definition on the sequence of the databases. Further, the professional database of the industry to which the problem belongs comprises a knowledge graph database, an SQL relational database and an RAG vector database, wherein the knowledge graph database is used for storing professional data of the industry to which the problem belongs. To solve the above technical problem, the present invention also provides a system for alleviating the illusion of a large model, which includes a processor, and the processor is configured to execute a computer program to implement the steps of the method for alleviating the illusion of a large model. The method for relieving the illusion of the large model has the advantages that the method comprises two stages of agents, wherein intention recognition analysis is firstly carried out on user problems through the first stage of agents, calling capability types corresponding to the user problems are selected, when the selected calling capability types are the calling of a search engine, the user problems are inquired through the search engine and query results are input into the large model through the first stage of agents, real-time knowledge information is acquired through the search engine, therefore timeliness of answering the problems of the large model can be guaranteed, the problem that the timeliness of answering the problems of the large model is poor due to the existing method using a vector library is solved, when the selected calling