CN-121996764-A - Vector node optimization method and system for enhancing RAG retrieval accuracy

Abstract

The invention discloses a vector node optimization method and a vector node optimization system for enhancing RAG retrieval accuracy, and particularly relates to the technical field of information retrieval and natural language processing, wherein the method comprises the steps of carrying out scene understanding and reconstruction on original query of a user to generate target query; based on target inquiry, calculating scene adaptation dynamic weights integrating semantic relativity, timeliness and authority for knowledge base document nodes, adaptively triggering multi-path searching according to scenes and complexity, optimizing preliminary sequencing by using the dynamic weights, performing multi-dimensional verification and reordering on primary results, outputting Gao Jiejie results, and continuously optimizing a system through feedback learning. The invention solves the problems of fuzzy query, static weight and single strategy in the traditional retrieval, realizes the spanning from semantic matching to scenerized precise matching, and remarkably improves the accuracy, adaptability and result reliability of the retrieval.

Inventors

• DENG WENPU
• Yin Wantao
• LEI JINLONG
• XIE XIN
• ZHANG SHIZHE
• Zhou Lingtian
• XIAO BANG
• GAO ZHENGBO

Assignees

• 华景乐游(深圳)智慧科技有限公司
• 郑州曙光云科技有限公司

Dates

Publication Date

20260508

Application Date

20260120

Claims (10)

1. 1. A vector node optimization method for enhancing RAG retrieval accuracy is characterized by comprising the following steps: Carrying out scene understanding and reconstruction on the original query input by the user to generate a target query; Generating dynamic weights which are matched with the scene and are used for vectorizing the document nodes by the knowledge base based on the target query; According to the result of scene understanding, self-adaptively triggering and executing multi-path searching, and obtaining a primary searching result set based on dynamic weight; And carrying out multidimensional verification and reordering on the primary retrieval result set, and outputting a high-order retrieval result.
2. 2. The method of claim 1, wherein the scene understanding and reconstruction comprises identifying a scene category to which the original query belongs based on text features of the original query and a user session context, and supplementing information and disambiguating the original query according to the scene category.
3. 3. The method of claim 1, wherein the dynamic weights of the scene adaptation are comprehensive weights fused with semantic relevance, timeliness and authority information, wherein the semantic relevance is determined based on vector similarity of the target query and the document nodes, the timeliness is dynamically determined based on document update time, and the authority is determined based on the document source level.
4. 4. The method of claim 1, wherein the adaptively triggering and executing the multi-path search comprises selecting at least one search strategy execution from a preset strategy set according to whether scene categories are professional fields and query complexity, wherein the preset strategy set comprises fast search based on vector similarity, mixed search combining keywords and vectors and enhanced search fusing knowledge maps.
5. 5. The method of claim 4, wherein when multiple search strategies are selected, the multiple searches are performed in parallel and the search results are fused and de-duplicated, and the multiple searches are fused with at least text vector similarity search and knowledge-graph-based entity relationship search.
6. 6. The method of claim 5, wherein the dynamic weights are used as core ranking factors in text vector similarity retrieval.
7. 7. The method of claim 1, wherein the multi-dimensional verification and reordering comprises calculating scores of documents in a primary search result set in a plurality of verification dimensions, weighting and fusing the scores of the dimensions with corresponding dynamic weights to obtain comprehensive confidence, and sorting and filtering out documents with low confidence according to the comprehensive confidence.
8. 8. The vector node optimization system for enhancing the RAG retrieval accuracy is characterized in that: for implementing the method of any one of claims 1-7, the system comprising: The semantic understanding and reconstructing module (1) is used for carrying out scene understanding and reconstructing on the original query input by the user to generate a target query; The dynamic weight calculation module (2) is used for calculating the dynamic weight of scene adaptation for the vectorized document nodes in the knowledge base based on the target query; the multi-path search execution module (3) is used for adaptively triggering and executing search according to the scene understanding result and obtaining a primary search result set based on dynamic weight; the fusion verification and sequencing module (4) is used for carrying out multidimensional verification and reordering on the primary retrieval result set; And the feedback learning module (5) is used for optimizing the semantic understanding and reconstruction module (1), the dynamic weight calculation module (2) or the multi-path retrieval execution module (3) according to the user interaction data.
9. 9. The system of claim 8, wherein the dynamic weight calculation module (2) forms the dynamic weights for scene adaptation by fusing semantic relevance weights based on vector similarity, timeliness weights based on update time, and authority weights based on source class.
10. 10. A computer-readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any one of claims 1-7.

Description

Vector node optimization method and system for enhancing RAG retrieval accuracy Technical Field The invention relates to the technical field of information retrieval and natural language processing, in particular to a vector node optimization method and a vector node optimization system for enhancing RAG retrieval accuracy. Background With the development of a large-scale pre-training language model, a retrieval enhancement generation RAG technology becomes a core paradigm for constructing a knowledge-intensive intelligent question-answering system, and the technology performs semantic retrieval in a vectorized knowledge base by converting user query into vectors and takes relevant documents as context input generation models so as to improve the accuracy and reliability of answers. However, existing RAG retrieval schemes still face significant challenges: Firstly, directly vectorizing an original query of a user, and effectively processing inherent ambiguity, information deletion and semantic ambiguity of the original query is difficult to achieve, so that a search intention source has deviation; Secondly, mainstream vector search generally only depends on static semantic similarity, and timeliness and authority of documents and dynamic adaptation relation between the documents and a specific application scene cannot be comprehensively considered, so that search results can be outdated, non-authoritative or mismatched with the scene; finally, the search strategy is often single and stiff, multiple search paths cannot be selected or fused in a self-adaptive manner according to the complexity and the field characteristics of the query, and the balance between efficiency and precision is difficult to achieve. These drawbacks limit the practical application effect and reliability of RAG systems in complex, professional and demanding scenarios. In view of the above, the invention provides a vector node optimization method and a vector node optimization system for enhancing RAG retrieval accuracy. Disclosure of Invention In order to overcome the above drawbacks of the prior art, the present invention provides a vector node optimization method and system for enhancing RAG retrieval accuracy, so as to solve the problems set forth in the background art. In order to achieve the above purpose, the invention provides a vector node optimization method for enhancing RAG retrieval accuracy, which comprises the following steps: Carrying out scene understanding and reconstruction on the original query input by the user to generate a target query; Generating dynamic weights which are matched with the scene and are used for vectorizing the document nodes by the knowledge base based on the target query; According to the result of scene understanding, self-adaptively triggering and executing multi-path searching, and obtaining a primary searching result set based on dynamic weight; And carrying out multidimensional verification and reordering on the primary retrieval result set, and outputting a high-order retrieval result. Preferably, the scene understanding and reconstructing specifically comprises the steps of identifying the scene category to which the original query belongs based on the text characteristics of the original query and the user session context, and carrying out information supplementation and disambiguation on the original query according to the scene category. Preferably, the dynamic weight of the scene adaptation is a comprehensive weight fused with semantic relativity, timeliness and authority information, wherein the semantic relativity is determined based on vector similarity between the target query and the document node, the timeliness is dynamically determined based on document update time, and the authority is determined based on the level of document sources. Preferably, the self-adaptive triggering and executing multi-path searching specifically comprises the steps of selecting at least one searching strategy from a preset strategy set to execute according to whether scene categories are professional fields and query complexity, wherein the preset strategy set comprises rapid searching based on vector similarity, mixed searching combining keywords and vectors and enhanced searching fusing knowledge maps. Preferably, when a plurality of search strategies are selected, the search results are executed in a parallel mode and are fused and de-duplicated, and the multi-path search is at least fused with the text vector similarity search and the entity relation search based on the knowledge graph. Preferably, in text vector similarity retrieval, dynamic weights are used as core ranking factors. Preferably, the multi-dimensional verification and reordering specifically comprises the steps of calculating the scores of all documents in a primary retrieval result set on a plurality of verification dimensions, carrying out weighted fusion on the scores of all the dimensions and corresponding dynamic weights t