Skip to content
Nightly retrieval eval, 90 days
image871 viewsAug 1, 2026

Nightly retrieval eval, 90 days

Sana Qureshi
Sana Qureshi@sanaqureshi

August 1, 2026 · 1 min read

871 views1 comment
image
Nightly retrieval eval, 90 days

Recall@5 over 90 nightly runs. The step down on day 38 is a corpus refresh nobody flagged; the recovery on day 51 is when we re-embedded. Three weeks of silent degradation that no code change would have explained.

Recall@5 over 90 nightly runs. The step down on day 38 is a corpus refresh nobody flagged; the recovery on day 51 is when we re-embedded. Three weeks of silent degradation that no code change would have explained.

Written by

Sana Qureshi

Sana Qureshi

ML engineer working on retrieval and inference, not training runs. I care about p99 latency, embedding drift, and whether the eval set actually resembles production traffic. Half my job is deleting models that were never better than the heuristic they replaced.

1 Comment

Sign in to join the discussion

Marcus Okafor
Marcus Okafor@marcusokafor1mo ago

Three weeks of silent degradation with no code change to blame is the entire argument for keeping a nightly metric nobody looks at. The value is all in the day you look back.

0