Embedding Model

Open Source Embedding Models (E5, BGE) challenge proprietary models like OpenAI's by offering state-of-the-art performance on retrieval benchmarks (MTEB) while being free to run locally.

Key Models

1. BGE (BAAI General Embedding)

• Performance: Consistently tops the MTEB leaderboard.
• Variants: available in large, base, and small sizes.
• Instruction-tuned: Requires specific prefix instructions for queries vs. passages.

2. E5 (Microsoft)

• Method: Text Embeddings by Weakly-Supervised Contrastive Pre-training.
• Quality: Strong performance on zero-shot retrieval tasks.
• Format: uses "query:" and "passage:" prefixes.

Comparison

• OpenAI Ada-002: Context length 8192, Pay-per-token, closed source.
• BGE-Large-en: Context length 512 (v1.5 supports longer), Free, Open Weights, Local privacy.

Use Cases

• Local RAG: Privacy-preserving document search without external APIs.
• Cost Reduction: Replacing paid embedding APIs for high-volume indexing.
• Custom Fine-tuning: Can be fine-tuned on domain-specific data (unlike closed APIs).