CRAG
Corrective Retrieval-Augmented Generation (CRAG) is an improved RAG paradigm that introduces a retrieval evaluator to score the relevance of retrieved documents. Based on those scores, the system decides whether to use the documents directly, discard them and trigger a web search, or refine the retrieved knowledge, thereby enabling self-correction of retrieval quality.
Core Facts
Timeline (last 90 days)
The advanced section covers CRAG, knowledge bases, Agent architecture, multi-Agent collaboration and tool calling, with hands-on development using mainstream frameworks.
CRAG adds an evaluation-and-correction mechanism on top of standard RAG: it first scores the relevance of retrieved documents, and if the score falls below a threshold, it triggers corrective steps such as web search or query rewriting.
CRAG (Corrective RAG) is an improved version of the standard RAG pipeline.
The project pushes pipeline progress to the frontend in real time via Server-Sent Events (SSE) and features a web dashboard with real-time visualization of the CRAG process.
CRAG's core innovation is the introduction of a lightweight retrieval evaluator that quantifies the relevance of each retrieved document as a confidence score and routes it into one of three processing paths accordingly: high scores are used directly, low scores are discarded and trigger a web search, and intermediate scores undergo knowledge refinement before use.
Query rewriting in CRAG has the LLM decompose the original question into several more precise sub-queries, retrieve results separately, and then merge them, so as to reduce retrieval failures caused by the vocabulary gap.
CRAG (Corrective Retrieval-Augmented Generation) evaluates and filters retrieval results for quality before generation, while Self-RAG has the model reflect on the reliability of citations during the generation process.
纠错式检索（CRAG）是2024年由Yan等人提出的自适应检索框架，引入检索评估器将检索结果分为正确、不正确、模糊三类并触发不同处理
All Facts (8)
CRAG adds an evaluation-and-correction mechanism on top of standard RAG: it first scores the relevance of retrieved documents, and if the score falls below a threshold, it triggers corrective steps such as web search or query rewriting.
65%UnverifiedThe advanced section covers CRAG, knowledge bases, Agent architecture, multi-Agent collaboration and tool calling, with hands-on development using mainstream frameworks.
50%UnverifiedCRAG (Corrective RAG) is an improved version of the standard RAG pipeline.
50%UnverifiedThe project pushes pipeline progress to the frontend in real time via Server-Sent Events (SSE) and features a web dashboard with real-time visualization of the CRAG process.
50%UnverifiedCRAG's core innovation is the introduction of a lightweight retrieval evaluator that quantifies the relevance of each retrieved document as a confidence score and routes it into one of three processing paths accordingly: high scores are used directly, low scores are discarded and trigger a web search, and intermediate scores undergo knowledge refinement before use.
50%UnverifiedQuery rewriting in CRAG has the LLM decompose the original question into several more precise sub-queries, retrieve results separately, and then merge them, so as to reduce retrieval failures caused by the vocabulary gap.
50%UnverifiedCRAG (Corrective Retrieval-Augmented Generation) evaluates and filters retrieval results for quality before generation, while Self-RAG has the model reflect on the reliability of citations during the generation process.
50%Unverified纠错式检索（CRAG）是2024年由Yan等人提出的自适应检索框架，引入检索评估器将检索结果分为正确、不正确、模糊三类并触发不同处理
50%