CN-122021880-A - Multi-step retrieval enhancement generation system optimization method based on inference chain evolution mechanism

CN122021880ACN 122021880 ACN122021880 ACN 122021880ACN-122021880-A

Abstract

The invention provides a multi-step retrieval enhancement generation system optimization method based on an inference chain evolution mechanism, which comprises the steps of (1) generating an initial inference chain according to an input problem, setting the input problem node as an unverified node, (2) globally identifying the unverified node, analyzing all nodes in the inference chain in each iteration, identifying the unverified node which obstructs the overall inference progress, (3) targeting retrieval and information integration, generating a corresponding retrieval query based on the unverified node, acquiring evidence documents related to the unverified node from an external knowledge base through a sparse and dense mixed retrieval mechanism, (4) carrying out fusion inference on the retrieved external evidence documents and the existing inference chain, generating a new inference chain, verifying and updating the node state, (5) carrying out inference termination and answer generation, wherein when all nodes in the inference chain are verified or reach preset iteration times, the large language model is combined with the inference chain to output a final result, and the method can realize global modeling and structuring control of the inference process, so that the system can obstruct the active identification of the multi-step key and the key retrieval problem can be solved.

Inventors

ZHANG PENG
WAN YUANLIANG

Assignees

天津大学

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (4)

1. The multi-step search enhancement generation system optimization method based on the inference chain evolution mechanism is characterized by comprising the following steps of: (1) The construction and initialization of the inference chain, namely generating an initial inference chain according to an input problem, and initializing the problem into an unverified node; (2) Global key node identification, namely carrying out global analysis on an inference chain in each iteration, and identifying unverified nodes which obstruct the overall inference progress; (3) Generating corresponding search queries based on unverified nodes, and acquiring evidence documents related to the unverified nodes from an external knowledge base through a sparse and dense mixed search mechanism; (4) The inference chain evolves and updates, namely, the relevant evidence document and the current inference chain are input into a large language model for fusion inference, a new inference chain is generated, and the node state is verified and updated; (5) And (3) reasoning termination and answer generation, namely outputting a final reasoning result by the system when all nodes are verified or reach the preset iteration times.
2. The multi-step search enhancement generation system optimization method based on an inference chain evolution mechanism of claim 1, wherein the initialization of the inference chain represents: constructing an initial inference chain based on input questions Wherein: the problem input by the user is a character string; Is a character string Representing initializing the node as an unverified node.
3. The multi-step search enhancement generation system optimization method based on an inference chain evolution mechanism of claim 2, wherein the inference chain Is represented by the structure: wherein Representing the chain of reasoning at step t, The j-th reasoning node of the t-th step is a character string type; representing the state of the jth node of step t, To enumerate string types, have And Two states, wherein the states are Is a fact node verified by a large language model, and the state is Is an unverified node judged by the large language model.
4. The method for optimizing a multi-step search enhancement generation system based on an inference chain evolution mechanism according to claim 3, wherein in the step 3, based on unverified nodes, a corresponding search query is generated through a large language model, and then a evidence document process related to the nodes is obtained from an external knowledge base through a sparse and dense mixed search mechanism, comprising: Will be The state of the nodes in the inference chain is The unverified nodes of the node are found one by one, and according to the original problem of the user and the context node of the node, the search query for the node is generated by utilizing a large language model heuristically, and the evidence documents related to the query are recorded as follows by a sparse and dense mixed search mechanism ; Using large language models to make them based on related evidence documents Heuristic will Conversion to the next inference chain Wherein the large language model is generated All node state labels in the document are obtained by a large language model according to related evidence documents And Is generated one by one.

Description

Multi-step retrieval enhancement generation system optimization method based on inference chain evolution mechanism Technical Field The invention belongs to the technical field of artificial intelligence and natural language processing, and particularly relates to a multi-step retrieval enhancement generation system optimization method based on an inference chain evolution system. Background With the rapid development of large language models, the large language models show strong language understanding and expression capability in tasks such as text generation, question-answering systems, knowledge extraction and the like. However, traditional large language models rely on internal parameter storage knowledge, and when faced with the problems of strong facts, high timeliness, or complex cross-document relevance, information loss and "illusion generation" phenomena often occur, resulting in inaccurate results or inconsistent logic. In order to solve the problem, the RAG technology introduces an external knowledge retrieval module in the generation process, so that the model can access an external knowledge base in real time, thereby improving the accuracy and the interpretability of the generated content. The existing RAG system is excellent in tasks such as open domain question answering, information extraction and knowledge auxiliary decision making, but most of the RAG systems adopt a single-step search and generation strategy, are difficult to deal with complex multi-hop reasoning tasks, and cannot establish logic connection among multiple documents or dynamically correct reasoning errors. To enhance reasoning capability, researchers have proposed multi-step adaptive search and reasoning methods that implement complex problem resolution and solution by gradually generating new search queries and updating contexts. However, such methods still have the problems of lack of global control of reasoning, difficulty in tracking a reasoning chain, and reasoning drift caused by information redundancy. Disclosure of Invention Aiming at the defects, the invention provides a multi-step RAG system based on an inference chain evolution mechanism, which effectively overcomes shortsighted decisions, error accumulation and logical inconsistency of a general multi-step RAG through structural modeling, global inference control and dynamic node optimization and provides a more robust, efficient and interpretable solution for complex inference tasks. The invention provides a multi-step search enhancement generation (RETRIEVAL-Augmented Generation, RAG) system optimization scheme based on an inference chain (Chain of Reasoning, coR) evolution mechanism, aiming at improving the inference accuracy of a search enhancement generation system in complex multi-hop reasoning, cross-document question-answering and knowledge-intensive tasks. The invention realizes global modeling and structural control of the reasoning process by constructing the 'reasoning chain' capable of dynamically evolving, so that the system can actively identify key nodes which obstruct problem solution in multi-step reasoning and conduct targeted retrieval and optimization. By the method, the model can be continuously focused on the unresolved sub-problems in the reasoning process, so that the problems of error accumulation and problem solving gravity center drifting in the traditional gradual reasoning mode are avoided, and the reasoning efficiency and result accuracy are remarkably improved. The core innovation of the invention is to provide a search enhancement generation method of 'inference chain evolution', which breaks through the limitation of the traditional single-step RAG and the disadvantages of the general multi-step RAG. The method leads the system to have the capabilities of self diagnosis, self correction and self optimization in the multi-step searching and generating process by introducing a structured global inference chain management mechanism. The invention is illustrated by the following technical scheme: the multi-step search enhancement generation system optimization method based on the inference chain evolution mechanism is characterized by comprising the following steps of: (1) The construction and initialization of the inference chain, namely generating an initial inference chain according to an input problem, and initializing the problem into an unverified node; (2) Global key node identification, namely carrying out global analysis on an inference chain in each iteration, and identifying unverified nodes which obstruct the overall inference progress; (3) Generating corresponding search queries based on unverified nodes, and acquiring evidence documents related to the unverified nodes from an external knowledge base through a sparse and dense mixed search mechanism; (4) The inference chain evolves and updates, namely, the relevant evidence document and the current inference chain are input into a large language model for fusion inf