RAG Answer Generation With Critique-Based Retrieval Gating

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing RAG-based generative AI systems face issues such as unnecessary resource consumption and time delays due to unnecessary searches, limited relevance evaluation between queries and search results, insufficient groundedness between search results and LLM answers, and lack of systematic evaluation of user queries and LLM answers, leading to inaccurate and unreliable answers.

Innovation Solution

A method and system that utilize a pretrained critique model to perform evaluation tasks, including determining whether to search for documents, evaluating relevance and groundedness, and assessing the utility of answers, to improve the performance of retrieval-augmented generation (RAG).

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If RAG-based generative AI system always refers to search results to generate answers, then answer accuracy and knowledge updating are improved, but resource consumption and time delay increase due to unnecessary searches

Engineering Contradiction:

Improveanswer accuracyVSAvoidresource consumption

Core Design Contradiction:

ReliabilityVSLoss of energy

Solution Approach 1:

The system performs preliminary evaluation of whether a search is necessary before actually executing the search. The critique model assesses the query and determines if retrieval-augmented generation is needed, preventing unnecessary search operations while maintaining the ability to provide accurate answers when search is beneficial

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system dynamically adjusts its behavior based on the query type and context. Instead of always performing searches, the system adapts its workflow by evaluating each query to determine the appropriate approach (search-based or direct generation), optimizing resource utilization while maintaining answer quality

Inventive Principle:

Principle #15Dynamics

2Reliability

If RAG-based generative AI system performs search for all queries, then knowledge-intensive task performance is improved, but time delay increases for tasks where search is unnecessary

Engineering Contradiction:

Improveknowledge updatingVSAvoidtime delay

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system autonomously evaluates each query to determine the appropriate processing path. The critique model self-assesses whether the query benefits from search-based approaches or can be handled through direct generation, enabling time-efficient processing without sacrificing knowledge accuracy

Inventive Principle:

Principle #25Self-service

3Device complexity

If RAG-based generative AI system lacks relevance evaluation between query and search result, then system complexity is reduced, but answer quality degrades due to inaccurate search results

Engineering Contradiction:

Improvesystem complexityVSAvoidanswer quality

Core Design Contradiction:

Device complexityVSReliability

Solution Approach 1:

The system incorporates feedback mechanisms where the critique model evaluates the relevance between queries and search results, and this evaluation feedback is used to adjust the answer generation process. This ensures that only relevant information is used to generate answers, maintaining high answer quality while managing system complexity through structured evaluation frameworks

Inventive Principle:

Principle #23Feedback

4Productivity

If RAG-based generative AI system lacks groundedness evaluation between search results and LLM answers, then generation speed is maintained, but answer reliability decreases due to mismatched information

Engineering Contradiction:

Improvegeneration speedVSAvoidanswer reliability

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system incorporates feedback mechanisms where the critique model evaluates the groundedness between queries, search results, and generated answers. This evaluation feedback is used to adjust the answer generation process, ensuring that generated answers are well-grounded in the retrieved information while maintaining efficient generation speeds through targeted verification

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS20260023769A1Method for generating answer based on advanced retrieval augmented generation and system therefor

Publication Date: 2026.01.22 SAMSUNG SDS CO LTD

AI summary

The disclosure relates to a high-performance RAG-based answer generation method, which includes: acquiring a query; performing a first evaluation task based on the query using a pre-trained critique model; retrieving documents related to the query based on a result of the first evaluation task; performing a second evaluation task based on the query and the retrieved documents using the critique model; and generating one or more answers, based on the query and one or more related documents, using a large language model (LLM) according to a result of the second evaluation task.