Compare · Concepts · Research notes
Embedding models — scored
Embedding choice fails when you mix models between index and query. Evaluate on your documents.
Last reviewed: 2026-08-15. Re-run the protocol after major model or pricing changes.
How do you compare embedding models?
- Build a 20-query set you already know the right chunk for.
- Embed with one model family for both index and query.
- Measure hit@5 and obvious mismatches.
- Only then consider dimension and price.
What is the scoring rubric?
| Criterion | Weight | How to score |
|---|---|---|
| Same model both sides | Gate | Fail mixed index/query models. |
| Hit@5 on your set | 1–5 | Your corpus beats a public leaderboard. |
| Hybrid need | Pass/Fail | Keywords still win on SKUs and IDs. |
| Ops cost | 1–5 | Re-embed cost when you change models. |
How should I choose an embedding model?
| Tool | Best fit | Start here | Watch-out | Note |
|---|---|---|---|---|
| Hosted embeddings | Fast pilot | Vendor default | Lock-in + re-embed cost | Fine for a first RAG. |
| Open embeddings | Local / portable | You own the vectors | You own evals too | Pair with the similarity lab. |
| Cohere Rerank | Second-stage rank | Boosts a shortlist | Not a replacement embedder | Use after retrieval, not instead of it. |
Learn the craft, not just the tool
Practice the transferable habit in What are embeddings?, then Embedding similarity lab, What is RAG?.
FAQ
What is the best embedding model?
The one that hits your 20-query set. Public leaderboards are a starting hint, not a purchase order.
Can I mix embedding models?
Not between the index and the query. You can add a reranker after retrieval.