Build your own RAG pipeline
Retrieve, measure, de-duplicate·5 min

Gold chunks & metrics

0 / 1

To measure retrieval you need a notion of relevant. Define it: a chunk is relevant to a question if its paragraphId is the question's goldParagraphId. The set of relevant chunk ids per question is its gold chunks — computed from the chunking, so it tracks whichever chunker you ran.

Two metrics over the gold chunks:

  • recall@k — the fraction of questions whose top-k retrieved chunks include at least one gold chunk. Did a gold chunk make the cut?
  • MRR@k — the mean of 1/rank of the first gold chunk in the top-k (0 if none). How high did it rank? (1-indexed.)
Worked example
query top-5 ids:  [ c9, c2(gold), c7, c1, c4 ]
  recall@5: a gold chunk is present       -> contributes 1
  MRR@5:    first gold chunk is at rank 2  -> reciprocal rank 1/2 = 0.5