Researchers at Salesforce AI Research introduced CodeXEmbed, a family of open-source embedding models specifically designed for code and text retrieval. These models, released in three sizes, SFR-Embedding-Code-400M_R, SFR-Embedding-Code-2B_R, and 7 billion parameters, address various programming languages and retrieval tasks. CodeXEmbed’s innovative training pipeline integrates 12 programming languages and transforms five distinct code retrieval categories into a unified framework. By supporting diverse tasks such as text-to-code, code-to-text, and hybrid retrievals, the model expands the boundaries of what retrieval systems can achieve, offering unprecedented flexibility and performance.
CodeXEmbed employs an innovative approach that transforms code-related tasks into a unified query-and-answer framework, enabling versatility across various scenarios. Text-to-code retrieval maps natural language queries to relevant code snippets, streamlining tasks like code generation and debugging. ......
