Source-free domain adaptive hashing enables cross-domain image retrieval without accessing source data. However, its reliance on extensive, costly target-domain annotations remains a major barrier. Addressing this limitation, we introduce the first framework uniting source-free and few-shot (1–5 shots) paradigms: Unified Elastic Prototype Contrastive Learning for domain-adaptive hashing (UEPCL). UEPCL achieves robust cross-domain retrieval with only minimal target supervision (1–5 images/class) and zero source data via three innovations: (1) Prototype-guided alignment constructs discriminative class prototypes from sparse targets to enforce intra-class compactness; (2) Contrastive hashing alignment maximizes inter-class separation while preserving semantic consistency; (3) Elastic knowledge distillation dynamically transfers source knowledge through probability matching, preventing catastrophic forgetting. Extensive experiments on benchmarks demonstrate UEPCL’s supremacy: it outperforms state-of-the-art methods by 2.34 to 8.70% mAP, 8.45% Top-50 precision and 12.11% Top-50 recall on average. This establishes a new frontier for efficient, privacy-compliant cross-domain retrieval. The code supporting this work will be made available at https://github.com/Luziji308/UEPCL.