netease-youdao/bcembedding is a package of embedding and reranking models for RAG released by NetEase Youdao. It is written in Python and follows the Apache-2.0 license. The repository has 1881 stars, and no publicly released version can be confirmed. The repository was created in January 2024.
Retrieval-augmented generation (RAG) is an approach that produces answers based on retrieved documents. BCEmbedding is introduced as a Chinese–English bilingual and cross-lingual RAG embedding project, providing both an embedding model that produces retrieval candidates and a reranking model that re-scores them. EmbeddingModel generates semantic vectors used for semantic search and question answering, while RerankerModel refines search results to assign ranks. Language support differs by model. EmbeddingModel supports Chinese and English, and RerankerModel supports Chinese, English, Japanese, and Korean. The README also states that it provides APIs for integration with LangChain and LlamaIndex.
There are limits worth noting. EmbeddingModel supports only Chinese and English, and Japanese and Korean fall within the supported scope of RerankerModel alone. Comparative performance benchmark figures are also not included in the information summarized here, so it is too early to judge how it compares with other embedding models. Judging by the documented scope alone, it is close to a bundle of embedding and reranking models tailored to Chinese–English-centric search.