BGE-EN-ICL

• 🧠 LLM-based English text embedding model from BAAI's BGE family.
• 🆕 Adds in-context learning: few-shot examples sharpen query embeddings.
• 🎯 Built for retrieval, clustering, and RAG vector search.
• 📏 Encodes queries and documents up to 512 tokens.
• 🔧 Runs via FlagEmbedding's in-context learning interface; optional FP16.
• 🔒 Released under the permissive Apache 2.0 license.
• 📚 Documented in the paper "Making Text Embedders Few-Shot Learners."

BAAI

The Beijing Academy of Artificial Intelligence (BAAI), also known as the Zhiyuan Institute, is a Chinese non-profit research laboratory dedicated to advancing the fundamentals of AI. Founded as a collaborative hub, BAAI brings together leading AI companies,…

BGE-EN-ICL is an English text-embedding model from the Beijing Academy of Artificial Intelligence (BAAI), part of the broader BGE (BAAI General Embedding) series. Unlike earlier BGE encoders, it is built on a large language model backbone and introduces in-context learning: by supplying a few task-relevant query–response examples in the prompt, the model encodes semantically richer queries and adapts to new tasks without fine-tuning. It produces a single dense embedding vector per input and is commonly used for retrieval, clustering, and other downstream tasks with vector databases.

Compared with its sibling BGE-M3, the two models target different needs. BGE-M3 is a multilingual, multi-functional model, while BGE-EN-ICL instead focuses on English, uses a 512-token limit for queries and documents, and leans on its in-context learning ability to boost task-specific representation quality.

Both models are released under the Apache 2.0 license, allowing personal and commercial use under the license terms. Implementation details, including the few-shot example format and last-token pooling, are documented in BAAI's model card and the accompanying paper. The model is straightforward to call through the FlagICLModel interface, where supplying or omitting task examples toggles the in-context learning behavior.

For teams already invested in the BGE ecosystem, BGE-EN-ICL offers an LLM-driven, example-conditioned alternative to the standard multilingual encoders, suited to English-centric retrieval and RAG pipelines where few-shot prompting can guide embedding behavior.