LAB · metrics-labRuns in your browser

Metrics Lab

Recall, precision, MRR and nDCG for keyword, vector and hybrid search over all 47 golden questions, computed in your browser, with the per-question ranks behind every number.

Used on RAG-39 and RAG-41 and RAG-44

Precision@k cannot pass about 1/k here: 46 of 47 questions have exactly one relevant chunk.

Retriever	Recall@3	Precision@3	MRR	nDCG@3

Per question

Retriever

Only questions outside the top 3

The last number is 1/rank: what this question adds to MRR.

What to notice

At k = 3, BM25 and hybrid reach recall 1.000. Switch to k = 1 and the order changes: rank 1 is a different contest.

Surprise

Precision@3 is about 0.34 for the best retriever. With one relevant chunk per question, that is near the maximum.

Try next

Pick vector search, tick “only questions outside the top k”, and diagnose each miss with the Retrieval Lab.