FlagOpen/FlagEmbedding

Retrieval and Retrieval-augmented LLMs

What it solves

BGE (BAAI General Embedding) provides a comprehensive toolkit for improving the retrieval stage of Retrieval-Augmented Generation (RAG) and search systems. It addresses the need for high-quality text and multimodal embeddings and reranking models to ensure that the most relevant documents are retrieved for an LLM to process.

How it works

The toolkit offers a suite of models and tools for the entire retrieval pipeline:

  • Embedding Models: Converts text or images into vectors (embeddings) for efficient similarity search. This includes specialized models like BGE-M3 for multi-lingual, multi-granular, and multi-functional retrieval (dense, lexical, and multi-vector).
  • Rerankers: Cross-encoder models that provide more accurate, though slower, scoring of the top-k documents returned by embedding models to refine the retrieval results.
  • Multimodal Capabilities: BGE-VL supports visual search applications, including text-to-image and image-to-text retrieval.
  • Fine-tuning & Evaluation: Includes tools to fine-tune embedders and rerankers on custom data and evaluate their performance using benchmarks like C-MTEB.

Who it’s for

This project is for developers and researchers building RAG-based applications, search engines, and multimodal retrieval systems who need state-of-the-art embedding and reranking capabilities.

Highlights

  • BGE-M3: A versatile model supporting 100+ languages, input lengths up to 8192 tokens, and three retrieval methods (dense, lexical, and multi-vector).
  • Multimodal Support: Includes BGE-VL for comprehensive visual search and Visualized-BGE for hybrid image-text data.
  • Comprehensive Toolkit: Provides a one-stop shop for inference, fine-tuning, and evaluation of both embedders and rerankers.
  • High Performance: Many models in the series rank highly on benchmarks like MTEB and C-MTEB.

Related

  • Project
    NovaSearch-Team/RAG-Retrieval

    A comprehensive toolkit for training, fine-tuning, and distilling RAG retrieval models, including embedding models, late interaction models, and rerankers.

  • Project
    embeddings-benchmark/mteb

    A multimodal toolbox for evaluating embeddings and retrieval systems, providing a standardized benchmark and leaderboard for comparing model performance.

  • Dispatch
    Qwen3-VL-Embedding and Qwen3-VL-Reranker Release

    Qwen has released Qwen3-VL-Embedding and Qwen3-VL-Reranker, a series of multimodal models designed for high-precision cross-modal retrieval across text, images, screenshots, and video.

  • Dispatch
    RAG Architectures: Avoiding Over-Engineering in Retrieval Augmented Generation

    Most RAG systems are over-engineered; starting with simple full-text search and query rewriting often solves 60% of use cases without the complexity of vector databases.

  • Project
    vespa-engine/sample-apps

    A collection of reference applications and tutorials for Vespa, demonstrating how to implement hybrid search, RAG, and multimodal vector retrieval at scale.