Embedding Models Compared¶
In one line: Pick an embedding model by language and content type, context length, vector size and your retrieval test results, not by a leaderboard rank, and plan for the fact that changing models means re-embedding everything. Last reviewed: 2026-11-07
The short version¶
- Embeddings turn text (or images) into vectors so you can search by meaning. Start with embeddings and vector search.
- Switching models re-embeds your whole corpus, so choose with a retrieval test on your own data first.
- Dimensions trade accuracy against storage and speed. Several vendors let you shorten vectors.
- Leaderboards (such as the MTEB leaderboard on Hugging Face) are a starting list. The page fetched for this review did not describe MTEB, so read its documentation yourself before relying on a score.
- Prices are not included; none were read.
Facts from vendor docs (as of 2026-11-07)¶
| Provider | Models named | Dimensions | Context / input | Notes |
|---|---|---|---|---|
| OpenAI | text-embedding-3-small, text-embedding-3-large, text-embedding-ada-002 | 1536, 3072, not stated | 8192 tokens | v3 models take a dimensions parameter; the page lists MTEB scores of 62.3%, 64.6% and 61.0%. If you shorten manually, normalise to unit length |
| Cohere | embed-v5.0-pro, embed-v5.0-fast, embed-v4.0, v3 English/multilingual and light variants | v5: 256 to 2048 (default 2048); v4: 256 to 1536; v3: 1024 or 384 | 128k tokens for v5 and v4; 512 for v3 | v5 and v4 handle text, images and mixed; v5 models share an embedding space so you can index with Pro and query with Fast; 100+ languages on v5 |
| Voyage AI | voyage-4-large, voyage-4, voyage-4-lite, voyage-code-4, voyage-finance-2, voyage-law-2; open-weight voyage-4-nano | 1024 default (256, 512, 2048 for the 4 series) | 32,000 tokens (16,000 for law-2) | Domain models for code, finance and law; voyage-4, voyage-4-large, voyage-4-lite and voyage-4-nano are compatible with each other (voyage-code-4 not stated) |
Open-source models (for example those on Hugging Face) are another route when you must run on your own hardware; none were fetched for this page.
How to choose¶
- List your needs: languages, text only or images too, code or legal content, self-hosting, data-residency rules.
- Build a small test set: 50 to 100 real questions with the document passages that answer them.
- Embed with two or three candidates, measure recall at the top 5 or 10 results, and compare.
- Test a shorter vector size if storage or latency matters.
- Record the model name and version with every stored vector so you know what needs re-embedding.
Common mistakes¶
- Picking the top leaderboard model without a test on your content.
- Mixing vectors from different models in one index. Only compatible families (Voyage 4 series except voyage-code-4 which the page does not state, Cohere v5 per their docs) can mix.
- Forgetting that queries and documents may need different input settings; check the vendor's reference.
- Chunking badly and blaming the model. See advanced RAG patterns.
Next steps¶
Sources¶
- OpenAI embeddings guide: https://developers.openai.com/api/docs/guides/embeddings (fetched 2026-11-07)
- Cohere Embed: https://docs.cohere.com/docs/cohere-embed (fetched 2026-11-07)
- Voyage AI embeddings: https://docs.voyageai.com/docs/embeddings (fetched 2026-11-07)
- MTEB leaderboard (Hugging Face Space): https://huggingface.co/spaces/mteb/leaderboard (fetched 2026-11-07; page gave no description)