CRAG

“existing approaches---including Self-RAG~asai2024selfrag and CRAG~yan2024crag---primarily target retrieval relevance without explicitly detecting or resolving contradictions”

“However, these methods still operate at the document level, failing to adequately filter individual text chunks.”

“it lacks the capability for high-level reasoning”

“CRAG provides a comprehensive benchmark with multi-hop and aggregation questions requiring cross-document synthesis, but operates on a fixed dataset without vertical-domain customization”

“CRAG~yan2024corrective, on the other hand, leverages the large-scale web search to supplement and rely on the vanilla LLM to integrate and refine knowledge from different sources. However, when the vanilla LLM fails to identify the defects in retrieved results, the whole pipeline would be broken and ineffective.”

“While these methods improve robustness against irrelevant context, they typically operate via Breadth-First Addition: they append new passages to the existing context.”

“CRAG~yan2024corrective uses an iterative approach with a small evaluator model but it still relies on GPT-3.5 for query rewriting.”

“While effective, these approaches often add supervision, special control tokens, auxiliary probers, or multi-stage loops that increase engineering complexity and latency.”

“While these approaches have improved robustness, leveraging LLMs' in-context learning capabilities in these scenarios is still underexplored.”

“Although CRAG improves the quality of retrieval, it does not address inaccuracies and irrelevancies in the final response.”

“the pre-processing methods introduce additional computational costs during inference and may lead to the loss of essential information.”