BGE-Large-EN v1.5

Text Tokenization

Converts input text into tokenized sequences using compatible tokenizers, preserving semantic structure for downstream embedding generation.

Transformer Encoding

Processes tokens through 24 transformer layers with self-attention to capture deep contextual relationships across the entire input.

Pooling Strategy

Applies mean pooling across token embeddings to produce fixed-length sentence vectors suitable for similarity and retrieval tasks.

Query Instructions

Supports optional instruction prefixes for queries (but not passages) to improve performance in asymmetric and instruction-aware retrieval scenarios.

Embedding Normalization

Normalizes output embeddings to unit length, enabling accurate cosine similarity comparisons between queries and documents.

Dense Retrieval

Maps text into high-dimensional dense vectors where semantic similarity corresponds to geometric proximity in embedding space.

Fine-Tuning Support

Allows task-specific fine-tuning while retaining strong zero-shot and general-purpose retrieval performance across NLP benchmarks.

Model Overview

• Model Name: BAAI/bge-large-en-v1.5
• Model Type: Dense text embedding model
• Language Support: English only
• Primary Use Cases: Semantic search, document retrieval, similarity comparison, clustering, reranking, NLP embedding tasks
• License: MIT / Open-source (distribution dependent)

Core Specifications

• Architecture: Encoder-only BERT-based embedding model (BGE v1.5 series)
• Parameter Count: ~335 million parameters
• Embedding Output Dimension: 1024-dimensional vectors
• Maximum Input Tokens: 512 tokens per input
• Model File Size: ~1.34 GB

Performance & Benchmarks

• MTEB Performance: Strong average results across tasks on the Massive Text Embedding Benchmark (MTEB)
• Similarity Distribution: Improved score distribution and retrieval effectiveness over earlier BGE v1 models

Functional Capabilities

• Text-to-vector embedding generation
• Semantic search and document retrieval
• Vector similarity scoring for ranking and clustering
• Support for cosine similarity and other vector metrics

Integration & APIs

• Hugging Face Transformers
• FlagEmbedding
• Sentence-Transformers
• LangChain and other LLM pipelines

Deployment Considerations

• Compute Requirements: GPU recommended for high-throughput inference; multi-core CPUs suitable for batch workloads
• Batching Support: Efficient batching for large-scale embedding generation
• Quantization Options: Third-party quantized formats (e.g., GGUF) available for reduced memory usage with minor accuracy trade-offs

Superior Embedding Quality

BGE-Large-EN v1.5 generates high-quality sentence embeddings optimized for accurate similarity calculations and semantic search tasks.

State-of-the-Art Retrieval

Achieves top performance on benchmarks such as MS MARCO and BEIR for dense retrieval and passage search applications.

Improved Similarity Distribution

Version 1.5 improves similarity score distribution, enabling more precise ranking and better separation of relevant and irrelevant documents.

Long Input Support

Handles extended text sequences effectively, making it suitable for longer queries and passages in retrieval workflows.

Flexible Fine-Tuning

Supports task-specific fine-tuning while maintaining strong zero-shot performance across diverse NLP use cases.

Semantic Search Optimized

Designed specifically for semantic search, question answering, and text classification with robust English language understanding.

Transformer Architecture

Built on an advanced transformer backbone with efficient self-attention mechanisms for contextual text representation.

Query Instruction Compatible

Supports query instructions for retrieval tasks, improving performance when distinguishing queries from passages.

Framework Integration

Compatible with Hugging Face Transformers, Sentence-Transformers, and FlagEmbedding for seamless deployment.

Production Ready

Optimized for real-world applications with normalized embeddings for cosine similarity and scalable inference pipelines.