BGE EN ICL
BAAI/bge-en-icl
published Jul 2024 · updated Jan 2025
BGE EN ICL is an embedding model that uses in-context learning with few-shot examples to produce high-quality text embeddings.
specs
| Task | Text Embedding |
| Architecture | Decoder-only LLM with in-context learning |
| License | MIT |
about this model
BAAI/bge-en-icl is an embedding model that leverages in-context learning (ICL) to produce high-quality text embeddings. By incorporating few-shot examples directly into the query input, the model adapts to new tasks without fine-tuning, generating embeddings that reflect the task structure defined by the provided examples.
Key Capabilities
The model integrates task-related examples into the query side, enabling it to handle both familiar and novel tasks through in-context learning. It retains the original decoder-only LLM framework, using last-token pooling, and supports both zero-shot and few-shot modes.
Benchmark Performance
BGE-EN-ICL achieves state-of-the-art results on the MTEB and AIR-Bench leaderboards.
BEIR (MTEB leaderboard):
AIR-Bench 24.04 — QA (nDCG@10):
| Model | wiki | web | news | healthcare | law | finance | arxiv | msmarco | ALL (8) |
|---|---|---|---|---|---|---|---|---|---|
| bge-en-icl zero-shot | 64.61 | 54.40 | 55.11 | 57.25 | 25.10 | 54.81 | 48.46 | 63.71 | 52.93 |
| bge-en-icl few-shot | 64.94 | 55.11 | 56.02 | 58.85 | 28.29 | 57.16 | 50.04 | 64.50 | 54.36 |
Long-Doc (en, Recall@10):
| Model | arxiv (4) | book (2) | healthcare (5) | law (4) | ALL (15) |
|---|---|---|---|---|---|
| text-embedding-3-large | 74.53 | 73.16 | 65.83 | 64.47 | 68.77 |
| e5-mistral-7b-instruct | 72.14 | 72.40 | — | — | — |
The model is released under the MIT license. The full training dataset (bge-full-data) contains over 2.1 million rows across 34 configs/splits, including sources such as arXiv, biorxiv, and newsgroups.
best for
- ·Retrieval-augmented generation (RAG) with few-shot examples
- ·Web search query to passage retrieval
- ·Domain-specific retrieval tasks (e.g., healthcare, law, finance)
FAQ
It uses in-context learning with few-shot examples in the query to produce high-quality text embeddings, achieving SOTA on MTEB and AIR-Bench.
It is released under the MIT license, free for both academic and commercial use.
Use the gigarouter OpenAI-compatible endpoint with your API key, sending queries and documents as input to get embeddings.
Queries should include a task instruction and optional few-shot examples, formatted with <instruct> and <query> tags. Documents are plain text.
The model card does not specify parameter count, but it is based on a decoder-only LLM and achieves SOTA on MTEB and AIR-Bench.
We're benchmarking and onboarding BGE EN ICL as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.