experiment · growing

Retrieval eval, in isolation

Hit rate, recall@k, MRR, NDCG. "Was the right thing retrieved" comes before answer quality.

Before measuring RAG end to end, measure it in isolation: for a given question, were the retrieved chunks the right ones? Answer quality depends on that answer.

Rendering diagram…

The dataset

Input is a question, output is the ground-truth document ids that answer it. Write 20 by hand, then grow it by generating synthetic questions from chunks.

MetricWhat it tells you
Hit rateis the right chunk in the top-k at all
recall@khow many of the relevant chunks came back
MRRhow high the first correct result sits
NDCGis the order right too — the metric for reranking

Doing RAG without a benchmark is turning knobs without knowing which change improved what.

#eval #mrr #ndcg

See this note on the whiteboard →