Embedding similarity metrics for RAG evaluation are deterministic and cheap but rigid because they reward matching the ground truth rather than actual correctness, and improvements can appear worse if the ground truth is narrow.

Authors

• Person: Sumit Umbardand Organization: LinkedIn
  RAG Hallucinations: Retrieval Success ≠ Generation Accuracy

Sources

• RAG Hallucinations: Retrieval Success ≠ Generation Accuracy www.linkedin.com Sumit Umbardand · LinkedIn via serper

Referenced by nodes (2)

• ground truth concept
• RAG evaluation concept