COSTA introduces a contrastive learning framework that enables effective code search across different programming languages without domain-specific training data.

• Uses unified contrastive learning to capture semantic similarities between natural language and code snippets
• Achieves SOTA performance for zero-shot cross-domain code search without expensive fine-tuning
• Employs domain-agnostic objectives that transfer effectively between programming languages
• Demonstrates robustness across multiple source and target language pairs

This advancement significantly reduces engineering costs for deploying code search systems across multiple programming languages, enabling more efficient software development workflows and knowledge transfer.

Zero-Shot Cross-Domain Code Search without Fine-Tuning