Search

CN-122018959-A - Gray scale publishing method, system, equipment and medium of large model knowledge base content

CN122018959ACN 122018959 ACN122018959 ACN 122018959ACN-122018959-A

Abstract

The application provides a gray level publishing method, a system, equipment and a medium of a large model knowledge base content, wherein the method comprises the steps of creating a gray level version of the knowledge base and setting a state of the knowledge base to be offline when gray level publishing is required, obtaining slice content to be gray level published, forming standard slice content after setting a gray level identification field on the standard slice content, storing the standard slice content in a vector database, reading the gray level version and the standard slice content of the knowledge base, adjusting the knowledge base based on the gray level version and the standard slice content of the knowledge base to form a gray level version knowledge base, correspondingly configuring the gray level identification field of the standard slice content, verifying the RAG effect of the gray level version knowledge base, and completing gray level publishing according to a verification result.

Inventors

  • DENG ZHI

Assignees

  • 上海数禾信息科技有限公司

Dates

Publication Date
20260512
Application Date
20251226

Claims (10)

  1. 1. A gray level distribution method of large model knowledge base content, the method comprising: When the gray level release of the knowledge base is needed, creating a gray level version of the knowledge base and setting the state as not on line; acquiring slice content to be subjected to gray release, forming standard slice content after setting a gray identification field on the standard slice content, and storing the standard slice content in a vector database; Reading the gray version of the knowledge base and the standard slice content, adjusting the knowledge base based on the gray version of the knowledge base and the standard slice content to form a gray version knowledge base, and correspondingly configuring the gray identification field of the standard slice content; And verifying the RAG effect of the gray version knowledge base, and finishing gray release according to the verification result.
  2. 2. The gray scale distribution method of contents of a large model knowledge base according to claim 1, wherein the gray scale identification field comprises a distributed released, a new add to gray scale and a remove to gray scale deletion.
  3. 3. The gray scale distribution method of large model knowledge base content according to claim 2, wherein said reading the knowledge base gray scale version and the standard slice content, adjusting the knowledge base based on the knowledge base gray scale version and the standard slice content to form a gray scale version knowledge base, and correspondingly configuring the gray scale identification field of the standard slice content, comprises: When the gray level version of the knowledge base corresponds to the newly added data, adding the standard slice content into the knowledge base to form the gray level version knowledge base, and setting the gray level identification field of the standard slice content as a new add to be gray level; When the gray version of the knowledge base corresponds to deleted data, deleting the standard slice content, and setting the gray identification field of the standard slice content as a remove to be deleted gray so as to form the gray version knowledge base; when the gray level version of the knowledge base corresponds to modification data, acquiring initial slice data before modification in the knowledge base, setting a gray level identification field of the initial slice data as a to-be-gray level deletion remove, newly adding the standard slice content to the knowledge base, and setting the gray level identification field of the standard slice content as a to-be-gray level newly added add to form the gray level version knowledge base.
  4. 4. A method of gray scale distribution of large model knowledge base content according to claim 3, wherein said verifying RAG effect of said gray scale version knowledge base comprises: When a common user of the knowledge base verifies the RAG effect, calling a formal RAG interface, and recalling the slice content of the knowledge base of a formal version through a first retrieval condition to verify the effect; and when the owner of the knowledge base verifies the RAG effect, calling a gray scale RAG interface, and recalling the slice content of the gray scale version knowledge base through a second retrieval condition to verify the RAG effect.
  5. 5. The gray scale distribution method of the large model knowledge base content according to claim 4, wherein the gray scale distribution is completed according to the verification result, comprising: When the verification effect meets the preset requirement, determining that the gray release verification is passed, setting the current formal version of the gray version knowledge base as a historical version, and setting the gray version of the knowledge base as an updated formal version; And modifying the gray scale identification field of the standard slice content with the gray scale identification field as the new add of the gray scale to be released released, and deleting the initial slice data with the gray scale identification field as the remove of the gray scale to be deleted.
  6. 6. The gray scale distribution method of the large model knowledge base content according to claim 4, wherein the gray scale distribution is completed according to the verification result, further comprising: When the verification effect does not meet the preset requirement, determining that the gray release verification is not passed, canceling the gray release, deleting the standard slice content of which the gray identification character is a new add to the gray to be detected, modifying the initial slice data of which the gray identification field is a remove to be detected to be released released, and deleting the gray version.
  7. 7. The gray scale distribution method of large model knowledge base contents according to claim 6, wherein it is determined that verification is not passed when the verification effect does not meet a preset requirement, and gray scale distribution is repeated after readjusting the standard slice contents based on the gray scale version.
  8. 8. A gray scale distribution system for large model knowledge base content, said system comprising: The gray version creation module is used for creating a gray version of the knowledge base and setting the state as not on-line when the knowledge base needs to be subjected to gray release; The content acquisition module is used for acquiring slice content to be subjected to gray release, forming standard slice content after setting a gray identification field on the standard slice content, and storing the standard slice content in a vector database; The version configuration module is used for reading the gray level version of the knowledge base and the standard slice content, adjusting the knowledge base based on the gray level version of the knowledge base and the standard slice content to form a gray level version knowledge base, and correspondingly configuring the gray level identification field of the standard slice content; And the verification and release module is used for verifying the RAG effect of the gray version knowledge base and completing gray release according to the verification result.
  9. 9. An electronic device is characterized by comprising a processor and a memory; the memory is used for storing a computer program; The processor is configured to execute the computer program stored in the memory, so that the electronic device executes the gray scale distribution method of the large model knowledge base content according to any one of claims 1 to 7.
  10. 10. A computer-readable storage medium having stored thereon a computer program, characterized in that the program, when executed, implements the gray scale distribution method of large model knowledge base contents according to any one of claims 1 to 7.

Description

Gray scale publishing method, system, equipment and medium of large model knowledge base content Technical Field The application belongs to the technical field of big data, relates to a knowledge base gray level publishing method, and particularly relates to a gray level publishing method, a system, equipment and a medium of a large model knowledge base content. Background The large language model (Large Language Model, LLM) refers to a deep learning model trained using large amounts of text data so that the model can generate natural language text or understand the meaning of language text. These models can provide in-depth knowledge and language production about various topics by training on a vast dataset. The key idea is to learn the mode and structure of natural language through large-scale unsupervised training to simulate the human language cognition and generation process to a certain extent. The knowledge base is a core component for providing external knowledge support for the large language model, and the essence of the knowledge base is the basis for dynamically supplementing the model to generate answers through a structured or unstructured data set. In the process of generating answers by using a large model, the accuracy of generating the answers is generally improved by adopting a retrieval enhancement generation technology, wherein the retrieval enhancement generation (RETRIEVAL-Augmented Generation, RAG) is an artificial intelligence technology framework combining information retrieval and text generation, and aims to improve the accuracy and reliability of a Large Language Model (LLM) in professional question and answer. ‌ its core is divided into two stages, ‌ retrieval stage ‌, in which the user questions are converted into vectors, related real-time information ‌ is retrieved from external knowledge base (such as database, document set), ‌ generation stage ‌, in which the retrieved results are input as context into LLM, and more accurate and well-defined answers are generated. In the prior art, the knowledge base used in the search enhancement generation technology generally does not support the release of the gray version of the change of the knowledge base, especially for the unverified change content, the direct effectiveness after the storage may cause adverse effect on the on-line RAG effect, and the last version cannot be rolled back when the RAG effect of the knowledge base is poor due to the change, thereby influencing the output result of the large model. Disclosure of Invention The application aims to provide a gray level release method, a system, equipment and a medium for large model knowledge base content, which are used for solving the problem that the output result of a large model is affected because the knowledge base cannot release gray level in the prior art. In a first aspect, the present application provides a gray level publishing method of a large model knowledge base content, the method comprising: When the gray level release of the knowledge base is needed, creating a gray level version of the knowledge base and setting the state as not on line; acquiring slice content to be subjected to gray release, forming standard slice content after setting a gray identification field on the standard slice content, and storing the standard slice content in a vector database; Reading the gray version of the knowledge base and the standard slice content, adjusting the knowledge base based on the gray version of the knowledge base and the standard slice content to form a gray version knowledge base, and correspondingly configuring the gray identification field of the standard slice content; And verifying the RAG effect of the gray version knowledge base, and finishing gray release according to the verification result. In an implementation manner of the first aspect, the gray identification field includes an issued released, a to-be-gray new add, and a to-be-gray delete remove. In an implementation manner of the first aspect, the reading the gray scale version of the knowledge base and the standard slice content, adjusting the knowledge base based on the gray scale version of the knowledge base and the standard slice content to form a gray scale version knowledge base, and correspondingly configuring the gray scale identification field of the standard slice content includes: When the gray level version of the knowledge base corresponds to the newly added data, adding the standard slice content into the knowledge base to form the gray level version knowledge base, and setting the gray level identification field of the standard slice content as a new add to be gray level; When the gray version of the knowledge base corresponds to deleted data, deleting the standard slice content, and setting the gray identification field of the standard slice content as a remove to be deleted gray so as to form the gray version knowledge base; when the gray level version of the knowledge base co