MachinoAI explainer / Retrieval-Augmented Generation
Corrective Retrieval Augmented Generation
Adds a retrieval-quality evaluator and corrective actions to make RAG robust when the initial retrieval is poor.
01
Abstract
PAPER-DERIVED FACTS: CRAG adds a retrieval evaluator to estimate whether retrieved documents are useful for a query. The evaluator routes the system toward direct generation, retrieval refinement, or supplementary web search depending on retrieval quality.
02
Introduction
PAPER-DERIVED FACTS: RAG systems often assume retrieved passages are relevant, but a poor retrieval set can cause the generator to amplify incorrect or irrelevant information. CRAG introduces an explicit correction mechanism.
03
Problem
PAPER-DERIVED FACTS: The system must detect when retrieval quality is insufficient and choose an appropriate recovery strategy. Static corpora may also lack the required evidence, motivating an external web-search fallback.
04
Background
PAPER-DERIVED FACTS: CRAG builds on conventional retrieve-then-generate RAG and adds retrieval evaluation before generation. It also introduces document refinement to remove irrelevant information from retrieved passages.
05
Methodology
PAPER-DERIVED FACTS: A lightweight evaluator assigns a confidence signal to retrieved documents. Depending on that signal, CRAG can use the retrieved set, refine its content, discard it and search the web, or combine corrected and external evidence.
Figure notes
Visual evidence
06
Architecture
PAPER-DERIVED FACTS: The pipeline contains retrieval, evaluation, routing, knowledge refinement, optional web search, and final generation. A decompose-then-recompose procedure extracts useful pieces from retrieved documents before they reach the generator.
07
Dataset
PAPER-DERIVED FACTS: Experiments cover four datasets spanning short-form and long-form generation tasks. The evaluation tests whether correction improves retrieval-augmented systems under varying retrieval quality.
08
Training
PAPER-DERIVED FACTS: The retrieval evaluator is trained separately from the generator and is designed to be lightweight. CRAG is presented as a plug-and-play correction layer that can be coupled with different RAG systems.
09
Experiments
PAPER-DERIVED FACTS: The authors compare corrected and uncorrected RAG configurations across multiple generation tasks. They examine how routing decisions and knowledge refinement affect final response quality.
10
Baselines
PAPER-DERIVED FACTS: Baselines include standard RAG systems without retrieval correction. The goal is to measure the incremental benefit of detecting poor retrieval and triggering corrective actions.
11
Results
PAPER-DERIVED FACTS: CRAG improves performance on the reported short- and long-form generation datasets. The authors attribute the gains to avoiding reliance on low-quality retrieved passages and supplementing missing information through web search.
12
Ablation
PAPER-DERIVED FACTS: The system components are evaluated through variants that remove retrieval evaluation, knowledge refinement, or web-search augmentation. These comparisons test whether each corrective action contributes to robustness.
13
Limitations
Error Analysis
PAPER-DERIVED FACTS: CRAG can still fail when the evaluator assigns an incorrect confidence score or when both the local corpus and web search lack reliable evidence. External search also introduces additional latency and source-quality concerns.
14
Conclusion
PAPER-DERIVED FACTS: CRAG makes retrieval quality an explicit control signal in RAG. Its main contribution is a lightweight correction layer that can recover from weak retrieval rather than passing every retrieved document directly to the generator.
15
References
PAPER-DERIVED FACTS: Primary source: Yan, S.-Q., Gu, J.-C., Zhu, Y., & Ling, Z.-H. (2024), Corrective Retrieval Augmented Generation. arXiv:2401.15884.
Continue reading
Related research
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Defines the canonical RAG architecture that combines a dense retriever, an external document index, and a seq2seq generator trained end-to-end.
Retrieval-Augmented GenerationDense Passage Retrieval for Open-Domain Question Answering
Introduces a practical dual-encoder dense retriever that replaces sparse lexical matching with learned semantic passage representations.
Retrieval-Augmented GenerationREALM
Introduces end-to-end retrieval during language-model pre-training so factual knowledge can live in an external, replaceable corpus rather than only in model parameters.