promptdojo_

Embeddings as features, not magic — step 2 of 7

Retrieval quality quietly degraded after last month's re-embedding job died halfway: half the corpus is embedded with the old model, half with the new one. Why is this so bad, and what's the fix?