Query Change Feedback for More Accurate RAG Result Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Generative large models in Retrieval-Augmented Generation (RAG) struggle to effectively utilize explicit feedback signals for training due to challenges such as quality and consistency of human feedback, scale and cost of feedback collection, timeliness of feedback, and data bias, which affect the relevance and accuracy of generated results.

Innovation Solution

A method that includes acquiring a change query corresponding to an input query, obtaining a reference result through search, and generating an output result using a generative model based on the input and change queries, with a training process that utilizes a prompt including the original query, change query, search result, and target instruction to improve relevance and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional search systems provide multiple results through recall, rough ranking, and precise ranking processes, then the system can meet basic search needs, but the relevance and accuracy of results to user intent are insufficient

Engineering Contradiction:
Improveaccuracy of generated resultVSAvoidcomplexity of search process
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent incorporates explicit feedback signals (change queries) from users into the generation process. By analyzing what users change in their queries when unsatisfied with initial results, the system learns to better align generated outputs with user intent, directly improving accuracy through feedback-driven optimization

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces change queries as an intermediary element that mediates between traditional search results and user satisfaction. These change queries serve as a bridge that captures user intent adjustments and translates them into training signals for the generation model, enabling more accurate result generation without completely redesigning the search pipeline

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If generative models are trained with human feedback to improve result quality, then the relevance of generated results improves, but the cost and scale of feedback collection increases

Engineering Contradiction:
Improverelevance of generated resultVSAvoidamount of feedback data
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The system enables self-service feedback collection by automatically extracting change queries from user search behavior without requiring explicit user input or manual annotation. Users naturally provide feedback through their query modification behavior, which the system automatically captures and utilizes for training, eliminating the need for costly manual feedback collection

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent uses change queries as copies or representations of user intent adjustments. Instead of collecting extensive raw feedback data, the system creates condensed representations (change queries) that capture the essential information needed for training, reducing the volume of data required while maintaining training effectiveness

Inventive Principle:
Principle #26Copying

3Measurement precision

If the system collects extensive feedback data to improve model training, then the model accuracy improves, but the timeliness of feedback utilization decreases

Engineering Contradiction:
Improveaccuracy of generated resultVSAvoidtimeliness of feedback
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary processing of change queries as they are generated during user search sessions. By immediately capturing and processing query changes in real-time rather than batching them for later processing, the system maintains feedback timeliness while accumulating sufficient training data

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements continuous feedback collection and model updating, where the system continuously learns from incoming change queries without interruption. This continuous action ensures that feedback is utilized in near-real-time, maintaining both accuracy improvement and timeliness by never stopping the learning process

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4668136A1Result generation method, generation model training method, device, and storage medium
Publication Date: 2025.12.24 BAIDU COM TIMES TECH (BEIJING) CO LTD
  • EP4668136A1 patent drawingFigure 1
  • EP4668136A1 patent drawingFigure 2
  • EP4668136A1 patent drawingFigure 3~4

AI summary

The disclosure provides a result generation method, a generation model training method, a device and a storage medium, relating to the field of computer technologies, and in particular, to the field of search and generative model technologies. The result generation method includes: acquiring (S301) a change query corresponding to an input query; obtaining (S302) a reference result by searching according to the input query and the change query corresponding to the input query; and generating (S303) an output result corresponding to the input query according to the input query, the change query corresponding to the input query and the reference result.