BAAI/bge-code-v1: A Generalist Code Embedding Model

BAAI/bge-code-v1 is a 1.5 billion parameter LLM-based embedding model developed by BAAI, specifically engineered for advanced retrieval tasks. This model stands out for its ability to handle code, text, and multilingual queries effectively.

Key Capabilities

• Superior Code Retrieval: Supports natural language queries in English and Chinese, alongside 20 programming languages, demonstrating exceptional performance in retrieving relevant code snippets.
• Robust Text Retrieval: Maintains strong capabilities comparable to dedicated text embedding models of similar scale.
• Extensive Multilingual Support: Excels in retrieval across multiple languages, including English, Chinese, Japanese, and French.

Performance Highlights

BGE-Code-v1 achieves state-of-the-art results on key benchmarks:

• CoIR Benchmark: Achieves an average score of 81.77, outperforming models like Voyage-Code-003 (78.53) and CodeXEmbed-7B (78.20).
• CodeRAG Benchmark: Scores an average of 72.8, surpassing SFR (67.0) and Jina-v2-code (65.4), with notable performance in tasks like DS-1000 (40.9) and SWE-bench-Lite (67.4).

Good For

• Code Search and Retrieval: Ideal for developers needing to find code based on natural language descriptions or code snippets.
• Text-to-Code Matching: Excellent for tasks like converting natural language questions into SQL queries or retrieving code implementations from textual explanations.
• Multilingual Applications: Suitable for projects requiring code and text retrieval across diverse language sets.