FlagEmbedding

Converts text input into fixed-dimensional dense vector representations using transformer-based encoder architectures (BGE v1/v1.5 models). Supports 100+ languages through unified embedding space training, enabling semantic similarity comparison across multilingual corpora. Implements contrastive learning with in-batch negatives and hard negative mining to optimize embedding quality for retrieval tasks.

BGE-M3 model generates three simultaneous embedding types per input: dense vectors (1024-dim), sparse vectors (lexical matching via learned vocabulary), and multi-vector representations (up to 8192 token context). Enables hybrid retrieval combining dense semantic search with sparse exact-match capabilities in a single forward pass, eliminating need for separate BM25 indexing.

Built-in evaluation system supporting BEIR (Benchmark for Information Retrieval) benchmark suite with 18 diverse retrieval tasks. Implements standard IR metrics (NDCG@10, MRR@10, MAP, Recall@k) and provides evaluation runners that handle data loading, retrieval execution, and metric computation. Enables reproducible model comparison and performance tracking across standard benchmarks.

Inference system supporting efficient batch processing of queries and documents with dynamic batching to maximize GPU utilization. Implements automatic batch size tuning, mixed-precision inference (FP16), and gradient checkpointing to reduce memory footprint. Supports both synchronous batch inference and asynchronous processing for high-throughput scenarios.

BGE-M3 and multilingual models enable cross-lingual retrieval by mapping queries and documents from different languages into unified embedding space. Supports retrieval across language boundaries without translation, enabling multilingual RAG systems. Implements language-agnostic dense and sparse representations learned through contrastive objectives on multilingual corpora.

BGE-ICL model enables embedding generation that adapts to task-specific contexts through in-context learning, allowing the embedding space to shift based on provided examples without fine-tuning. Implements prompt-based adaptation where query and document embeddings are influenced by demonstration examples, enabling zero-shot task transfer for domain-specific retrieval.

Base reranker models (BGE-reranker-large, BGE-reranker-base) implement cross-encoder architecture that scores document-query pairs directly by processing both inputs jointly through a transformer, producing relevance scores. Unlike embedding-based retrieval, rerankers see full context of both query and document, enabling more accurate ranking but at higher computational cost. Typically applied as second-stage ranker after initial retrieval.

BGE-reranker-v2-gemma and similar LLM rerankers use decoder-only language models to generate relevance scores or explanations for document-query pairs. Instead of classification-based scoring, these models generate tokens representing relevance (e.g., 'Yes', 'No', or numeric scores), leveraging LLM reasoning capabilities for more nuanced ranking decisions. Enables interpretable reranking with optional explanation generation.

FlagEmbedding provides layerwise and lightweight reranker variants optimizing for different deployment constraints. Layerwise rerankers use intermediate layer outputs for faster scoring with minimal accuracy loss. Lightweight variants use smaller model architectures (MiniCPM-based) reducing memory footprint and inference latency while maintaining reasonable ranking quality. Enables deployment on resource-constrained environments.

FlagEmbedding provides end-to-end fine-tuning infrastructure for both embedder and reranker models on custom datasets. Implements contrastive learning objectives for embedders (in-batch negatives, hard negative mining) and ranking losses for rerankers (pairwise ranking, listwise losses). Includes data preparation utilities, training loops with distributed support, and evaluation metrics to measure fine-tuning effectiveness.

Utility module that identifies hard negatives (documents that are semantically similar to query but not relevant) from unlabeled corpora using initial embeddings. Augments training datasets by replacing random negatives with hard negatives, improving model robustness to false positives. Implements efficient batch processing to mine negatives from large corpora without exhaustive comparison.

Framework for distilling large embedding and reranker models into smaller student models, preserving performance while reducing inference latency and memory footprint. Uses teacher-student training where student model learns to match teacher embeddings or scores through KL divergence or MSE losses. Enables deployment of high-quality models on resource-constrained devices.

Auto model loading system that automatically detects model type (embedder vs reranker, encoder-only vs decoder-only, base vs specialized) and instantiates appropriate inference class. Abstracts away architecture-specific implementation details, providing unified interface for loading any BGE model variant. Supports loading from Hugging Face Hub or local paths with automatic configuration parsing.