(논문 요약) Searching for Best Practices in Retrieval-Augmented Generation (Paper)

핵심 내용

• RAG 의 부분들을 ablation test
  
• Query classification: BERT-base-multilingual 로 다음 15 class classification
  
• Embedding model 비교
  
• Vector database
  
  • Multiple index types: flexibility in search optimization
  • Billion-scale: handle large datasets
  • Hybrid search: vector search + traditional keyword search
  • Cloud-native: cloud environments
• Retrieval methods
  • Query Rewriting
  • Query Decomposition
  • Pseudo-documents Generation: query 를 바탕으로 hypothetical document 를 생성하여, text embedding matching 시 사용
• Reranking (ordering)
  • Deep Learning Models
  • TILDE Reranking (이라는 방법…)

실험 결과