Topics

In Anthropic’s article on Contextual RAG, they ran experiments and give these tips:

  • Generate chunks with context
  • For embeddings, use Gemini or Voyage
  • Combine embeddings with BM25 for keyword matching which is one of the common vanilla RAG problems
  • Add reranking stage after retrieval to reduce the failure rate by 67% (5.7% to 1.9%)
    • Use top-K with K=20 and top-N with N=150