A reproducible test of BM25, embeddings, and parent documents shows why author-written evals overrate search, and why answer position in context matters.
The post RAG Retrieval Lost 40 Points When Distracting Passages Grew to 2.68M appeared first on Towards Data Science.
