Compare · Concepts
Embedding models compared
Embeddings power semantic search and RAG. Model choice matters — but chunking, metadata filters, and hybrid keyword search often matter more.
Selection checklist
- Same model at index time and query time
- Language and domain coverage for your corpus
- Latency and cost per 1M tokens
- License (especially for commercial local models)
- Eval on your questions — not only public MTEB screenshots
Open vs hosted
Hosted APIs are easy to start and scale. Open/local embeddings help with privacy and cost at volume. Learn the concept in What are embeddings? and practice selection in pick an embedding model.
Practical default
Start with a strong general embedding + hybrid search. Add reranking if top-k looks noisy. See also embedding and vector database.
FAQ
Can I mix embedding models in one index?
No. Query embeddings must come from the same model (and usually version) used at index time. Re-embed when you switch.
Are higher dimensions always better?
Not always. Higher dimensions can help recall but cost more storage and latency. Evaluate on your corpus with hybrid search as a baseline.