BGE
A toolkit of open embedding and reranking models for search and retrieval-augmented generation (RAG).FlagOpen/FlagEmbedding
- Stars
- 12.3k
- Forks
- 945
- Licence
- MIT
- Language
- Python
- Last update
- 10 days ago
- Type
- Library
What it is
FlagEmbedding (the BGE project) provides text embedding models that turn words and passages into numeric vectors, so software can find text with similar meaning. It also includes reranker models that re-order search results, and tools for fine-tuning these models. People use it to build semantic search and to supply relevant documents to large language models in RAG systems.
- Encode a document collection into vectors and search it by meaning
- Re-rank the top results from a first-stage search with a reranker model
- Feed retrieved passages into an LLM to answer questions over private text
- Fine-tune an embedding model on your own query and passage pairs
Great for
- Developers building semantic search over their own documents
- Teams adding retrieval-augmented generation (RAG) to an LLM app
- Researchers comparing or fine-tuning embedding and reranker models
- Multilingual projects that need one embedding model for many languages
Not ideal for
- People who want a finished search app without writing code, since this is a toolkit of models and scripts
- Projects that cannot run machine-learning models locally or pay for hosted inference, since the models need compute
- Anyone who needs guaranteed stable APIs, as the README lists many research projects and frequent new releases
How to get it
- 1
Install with a package manager
You'll need: Python
Clone the repository
git clone https://github.com/FlagOpen/FlagEmbedding - 2
Run it
- Install the package as described in the README's installation section
- Load a BGE embedding model from Hugging Face
- Encode a small set of sentences and compare their similarity
- Optionally add a reranker to order the top search results
- 3
Get the most out of it
- Use a reranker on top of embedding results to improve ranking quality
- Follow the tutorials folder for worked examples
- Check the BGE documentation site linked from the README for model details
- Use the Hugging Face model collection to pick a model for your language or task
Downloads
Strengths
- MIT licensed, so the code can be used commercially according to the README
- Broad set of models, including multilingual and multimodal (BGE-VL) variants
- Documentation and tutorials are maintained alongside the code
Trade-offs
- Many components are research projects with varying levels of documentation
- Larger models need substantial GPU or memory resources
- The README is long and spread across many sub-projects, which takes time to navigate
How it compares
| Project | Stars | Type | Level | Updated | Best for |
|---|---|---|---|---|---|
| FlagEmbedding (this) | 12.3k | Library | intermediate | 10 days ago | Developers building semantic search over their own documents |
| llama_index | 52.8k | Library | intermediate | 10 days ago | Python developers building retrieval-augmented generation (RAG) apps over private documents |
| PageIndex | 40k | Library | intermediate | 7 days ago | Developers building question-answering over long PDFs such as annual reports or manuals |
| LightRAG | 40.5k | Library | intermediate | 7 days ago | Developers building question-answering over their own document collections |
| RAG-Anything | 24.4k | Library | intermediate | 4 days ago | Developers building question-answering systems over PDFs or Office documents with figures and tables |
| turbovec | 17.5k | Library | intermediate | 30 days ago | Python or Rust developers building semantic search over a few million or fewer embeddings on a single machine |
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.