Embedding模型排行榜2026：怎么给RAG选对Embedding（MTEB指南）
换了Embedding模型后RAG质量变了，却说不清原因。Embedding是检索质量的隐形推手，选择需要方法论。
Embedding Models: The Most Underrated Decision in RAG
Retrieval quality is mostly embedding quality. The model that converts text into vectors decides which chunks are near each other - and every downstream answer inherits that decision. The standard reference for comparing embeddings is MTEB (Massive Text Embedding Benchmark), a community benchmark maintained on Hugging Face covering dozens of tasks across retrieval, clustering, classification, and multilingual datasets.
What MTEB Actually Tells You
MTEB scores models on retrieval (how well they find relevant documents), reranking, and other tasks. But the leaderboard average hides what matters: a model ranked #1 overall can be mediocre on your specific domain. Medical text, code, Chinese-language content, and short product descriptions all favor different models. The leaderboard is a starting point, never a conclusion.
The Selection Method That Works
- Match the task. For RAG, filter MTEB by the retrieval task and by language - if your corpus is Chinese or mixed, look at multilingual columns, not the English ranking.
- Test on your data. Build a 200-500 question set from your real corpus and measure retrieval hit-rate per model. This 30-minute experiment is worth more than any leaderboard.
- Consider the dimension trade-off. Embedding dimension affects storage and speed: 384-dim models (like bge-small) are fast and cheap; 1024-3072-dim models (bge-m3, modern flagship models) retrieve better but cost more RAM and latency.
- Check context length. Older models truncate at 512 tokens; newer ones handle 8K. Long documents need the latter.
Practical Recommendations for 2026
- Multilingual needs: bge-m3 and similar multilingual models are the default choice for Chinese+English corpora.
- English-only, cost-sensitive: small models (bge-small, e5-small) run on CPU and cover most cases.
- Maximum quality: the top MTEB open models (bge-m3 class, or proprietary embeddings from OpenAI/Cohere if budget allows).
- Domain-specific: fine-tune an open embedding model on 1-5k labeled pairs - it beats switching to a bigger generic model.
FAQ
Is MTEB free? Yes - it is an open benchmark on Hugging Face.
Can I run embeddings locally? Yes - most open models run on CPU, which is why self-hosted RAG is cheap.
How often should I re-evaluate my embedding choice? Whenever your corpus composition changes materially, or every few months as new models land.
Bigger embedding model always better? No - cost, latency, and dimension matter; match the model to your data, not to the top of the leaderboard.
