Qdrant experiments: 10→500 candidate depth lifts best-possible nDCG by 0.28 but real score by ≤0.01

qdrant_engine · x · 2026-09-21

Qdrant tested candidate depth across five public datasets: raising the limit from 10 to 500 improved the best-possible nDCG@10 by up to 0.28, yet the actual pipeline score changed by at most 0.01—documents are often retrieved, just not ranked highly enough. Practical guidance: start testing at limit 100 and 200 for reranker stages, compare approximate recall against exact search before raising hnswef, and try quantization before cutting depth when RAM-bound.

Original post →

More from coding & agent

coding & agent channel →