FlagOpen/FlagEmbedding provides a comprehensive toolkit for retrieval-augmented LLMs, offering a suite of embedding and reranking models. It targets researchers and developers building search and RAG systems, enabling state-of-the-art performance across various languages and retrieval tasks.

How It Works

The project leverages advanced transformer architectures, including LLM-based models, to generate dense embeddings. It supports multiple retrieval paradigms such as dense, lexical, and multi-vector (ColBERT) retrieval, unifying these functionalities within single models like BGE-M3. This multi-faceted approach enhances retrieval accuracy and flexibility.

Quick Start & Requirements

• Installation: pip install -U FlagEmbedding (for inference) or pip install -U FlagEmbedding[finetune] (for fine-tuning).
• Dependencies: Python. GPU with CUDA is recommended for optimal performance.
• Usage: Load models via FlagAutoModel.from_finetuned and use the .encode() method. See embedder inference and reranker inference for details.

Highlighted Details

• BGE-M3 offers multi-linguality (100+ languages), multi-granularity (up to 8192 tokens), and multi-functionality (dense, lexical, multi-vec retrieval).
• BGE-VL provides state-of-the-art multimodal embedding capabilities for diverse visual search applications.
• The toolkit includes lightweight rerankers and models fine-tuned with techniques like LM-Cocktail for improved resilience.
• Active development includes new benchmarks like MLVU for long video understanding and AIR-Bench for fair OOD evaluation.

Maintenance & Community

The project is actively maintained with frequent updates and new model releases. Community engagement is encouraged via WeChat groups. Tutorials are continuously updated.

Licensing & Compatibility

FlagEmbedding is licensed under the MIT License, permitting both academic and commercial use without significant restrictions.

Limitations & Caveats

While the project offers extensive multilingual support, specific performance nuances may exist across all languages. Some newer models like BGE-VL are released under MIT, but the README also mentions other projects with potentially different licenses, requiring careful verification for specific components.