BCEmbedding is an open-source model family on GitHub from NetEase Youdao, providing bilingual and cross-lingual embedding and reranker models built specifically for retrieval-augmented generation. The repository has about 1,900 stars and remains actively maintained, with updates continuing into 2026.
RAG systems live or die by retrieval quality, and BCEmbedding attacks both stages of the problem. Its embedding model, bce-embedding-base_v1, maps Chinese and English text into vectors that work across languages, so a query in one language can retrieve documents in the other. Its companion reranker, bce-reranker-base_v1, then re-scores the retrieved candidates to push the most relevant passages to the top. Both models posted strong results on public benchmarks such as MTEB and C-MTEB at release, and the repository provides easy integration paths, including LangChain support and deployment guidance. For teams building Chinese-English knowledge bases, these models are a proven, production-friendly choice.
The models use a compact base-model size that keeps inference fast and cheap on a single GPU while still delivering retrieval quality competitive with much larger alternatives. The repository documents evaluation methodology in detail and provides both embedding-only and combined retrieval-plus-reranking pipelines, so teams can adopt the full stack or just the stage they are missing in an existing search system.
BCEmbedding is free and open source under the Apache-2.0 license. The models can be downloaded and self-hosted without license fees, API charges or per-query costs from NetEase Youdao.
BCEmbedding is for RAG developers, search engineers and AI product teams, especially those serving Chinese and English users, who need dependable open embedding and reranking models they can run on their own infrastructure.