FlagEmbedding

Overview

FlagEmbedding is a comprehensive retrieval and retrieval-augmented LLM framework developed by the Beijing Academy of Artificial Intelligence (BAAI). It includes the BGE (BAAI General Embedding) series of state-of-the-art embedding models.

BGE Model Series

BGE v1.5 (bge-*-v1.5)

Improved version addressing similarity distribution issues:

• Enhanced retrieval ability without requiring instructions
• Available in large, base, and small sizes
• Top performance on MTEB and C-MTEB benchmarks

BGE-M3 (Multi-Functionality, Multi-Linguality, Multi-Granularity)

The flagship model with unique versatility:

Multi-Functionality

• Dense Retrieval: Traditional vector similarity search
• Multi-Vector Retrieval: Multiple vector representations per document
• Sparse Retrieval: Keyword-based retrieval like BM25
• Supports all three retrieval modes simultaneously

Multi-Linguality

• Supports over 100 languages
• Trained on balanced multilingual datasets
• Strong cross-lingual capabilities

Multi-Granularity

• Short Texts: Single sentences
• Medium Texts: Paragraphs
• Long Documents: Up to 8,192 tokens
• Handles various input lengths effectively

Training Methodology

• Pre-training: RetroMAE approach
• Fine-tuning: Large-scale pairs data with contrastive learning
• Data Quality: Curated high-quality training datasets

Additional Models

Reranker Models

• bge-reranker-base: Cross-encoder for reranking
• bge-reranker-large: Larger, more powerful reranker
• bge-reranker-v2-m3: Latest reranker with multilingual support
• More accurate than embedding-only approaches
• Recommended for re-ranking top-k retrieved documents

Performance

• MTEB Leaderboard: Ranked #1 for English embeddings
• C-MTEB: Top performance on Chinese benchmark
• Multilingual Tasks: Strong performance across 100+ languages
• Retrieval Quality: Superior recall and precision metrics

Key Features

• Open Source: Fully open-source under permissive license
• Easy Integration: Compatible with popular frameworks
• Fine-tuning Support: Can be adapted to specific domains
• Production Ready: Battle-tested in real-world applications

Use Cases

• Semantic search across languages
• RAG (Retrieval-Augmented Generation) systems
• Document retrieval and ranking
• Clustering and classification
• Cross-lingual information retrieval
• Question answering systems
• Recommendation engines

Integration

Framework Support

• LangChain
• LlamaIndex
• Haystack
• HuggingFace Transformers
• Sentence Transformers

Deployment Options

• Local inference
• Cloud APIs
• Together AI platform
• Amazon Bedrock (fine-tuning support)

Model Sizes

• Large: Maximum accuracy, higher compute
• Base: Balanced performance and efficiency
• Small: Lightweight, faster inference

Technical Details

• Architecture: Transformer-based encoders
• Context Window: Up to 8,192 tokens (BGE-M3)
• Embedding Dimensions: Model-dependent (typically 768-1024)
• Batch Processing: Optimized for throughput

Pricing

Free and open-source. Available on HuggingFace Hub for self-hosting or through commercial API providers.