Corrective Retrieval Augmented Generation

Meta info.

• Authors: Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling
• Paper: https://arxiv.org/pdf/2401.15884.pdf
• Affiliation: Google Research, UCLA, USTC
• Code: https://github.com/HuskyInSalt/CRAG

TL; DR

confidence score, web search, knowledge refinement로 잘못 찾아온, 혹은 최적이 아닌 결과를 self-correction하여 모델 생성 결과에 hallucination 감소

Sggestions

• 간단한 evaluator 모델이 retrieval 결과의 relevance score 계산 (T5-large)
• Action Trigger가 3-level로 confidence degree 구분
  • correct → (internal_knowledge) refinement operation (knowledge decomposition-then-recomposition)
  • incorrect → (external_knowledge) discard and then web search (chatgpt가 검색어 꼴로 rewrite)
  • ambiguous → (both) combine retrieved results and web search results
• 임의의 generative model이 생성