Embedding Fine-Tuning When to Fine-Tune | Signal | Action | |---|---| | Domain vocabulary very different from web (legal, medical, code) | Fine-tune | | Generic model recall@10 above 0.85 on eval | Do not fine-tune; tune retrieval | | No labelled (query, doc) pairs available | Start with GPL synthetic queries | | Fewer than 500 real pairs | Use synthetic + in-batch negatives | | 10k+ labelled pairs with hard negatives | Expect 5-15% nDCG lift | Contrastive Loss Choices | Loss | Inputs | When to use | |---|---|---| | MultipleNegativesRankingLoss (MNR) | (anchor, positive) | Default — uses in-b…