Embedding models for RAG have converged (agentset.ai)

A comparative evaluation of 13 embedding models (OpenAI, Cohere, Google, Voyage, Qwen, Jina, BAAI, BGE, Gemini, etc.) across 8 diverse datasets (essays, financial QA, business docs, multilingual QA) used pairwise retrievals judged by an LLM (ChatGPT 5) and aggregated with ELO scoring. The headline: embeddings have largely converged. Seven of 13 models meet or exceed the 1500 baseline, 11 of 13 lie within a 50‑point ELO band (about 85% clustered), the top 4 are separated by only ~23.5 ELO points, and rank‑1 vs rank‑10 differs by roughly 3%. Only two clear underperformers emerged (Qwen3‑0.6B and Gemini‑004), and even the top model wins only ~56% of head‑to‑head matchups. Why it matters: differences that look big on paper shrink in realistic retrieval settings, so picking an embedding is now more about cost, latency, and deployment than raw retrieval accuracy. The study finds no consistent correlation between model size, price, or latency and embedding quality. Technically, convergence is expected—models trained on similar data with similar objectives map semantics into vector spaces that quickly reach diminishing returns. For RAG systems, the biggest practical gains will come from engineering—chunking strategy, hybrid search, indexing, and reranking—rather than swapping embedding providers. Full rankings are available on the embedding leaderboard.