Embedding models: bge, E5, and NV-Embeddings compared

Observations on retrieval and classification: - bge-large: Strong general retrieval, good multilingual variants; higher memory. - E5-mistral: Competitive with smaller footprint; efficient for CPU serving. - NV-Embeddings: Optimized for NVIDIA stack; great throughput on GPUs. Takeaways: - Normalize embeddings and tune chunk sizes; these matter more than model choice at times. - Use task-specific embeddings for code or multilingual content. - Add a reranker; expect 20-40% boost in answer quality. Verdict: Start with bge for baseline, E5 if CPU-bound, NV if you have GPU infra and need speed.