I'm building a RAG app and need to serve an embedding model, what are good options?

Good options depend mainly on your constraints: quality, latency, cost, language coverage, and whether you want to self-host or use an API. For a RAG app, these are common choices:

Strong hosted/API options

1) OpenAI text-embedding-3-small / text-embedding-3-large

• Very strong general-purpose embeddings
• Easy to use, reliable, fast
• Good default choice if API use is fine
• 3-small is usually the cost-effective pick; 3-large gives better retrieval quality

2) Cohere Embed

• Strong retrieval-oriented embeddings
• Good multilingual performance
• Often a solid choice for enterprise/search use cases

3) Voyage AI embeddings

• Known for strong retrieval quality
• Good option if you want top-tier embedding performance

Strong self-hosted/open models

4) BAAI bge family

• Popular open-source choice for RAG
• Good quality, widely used
• Examples: bge-small, bge-base, bge-large, and newer variants like bge-m3 for multilingual/multi-function use

5) e5 family

• Another widely used open embedding family
• Good for semantic search / retrieval
• Works well when fine-tuned or used with the recommended query/document prefixing patterns

6) Jina embeddings

• Good open/self-hostable options
• Often competitive for retrieval and multilingual tasks

7) SentenceTransformers models

• Huge ecosystem and lots of model choices
• Good if you want flexibility and easy local deployment
• Common starting points include models from all-MiniLM family for lighter setups, though quality is lower than newer retrieval-focused models

What I’d recommend in practice

If you want the easiest high-quality setup:

• OpenAI text-embedding-3-small to start
• Switch to text-embedding-3-large if retrieval quality matters more than cost

If you want to self-host:

• bge-m3 is a very good modern default
• e5-large or bge-large are also solid

How to choose

Consider:

• Quality: test on your own queries and docs
• Latency: smaller models are faster
• Cost: hosted APIs cost per token/vector; self-hosting costs infra
• Multilingual support: bge-m3, Cohere, Voyage, and some multilingual models are strong
• Domain fit: legal, code, support tickets, and technical docs may benefit from specific evaluation

Practical advice

For RAG, don’t just pick one blindly:

1. Build a small eval set of real queries
2. Measure retrieval metrics like recall@k / nDCG
3. Test chunking + embedding model together
4. If possible, compare:
   • text-embedding-3-small
   • text-embedding-3-large
   • bge-m3 or e5-large

If you want, I can give you:

• a short list by budget
• a self-hosted deployment stack
• or a benchmark plan for choosing embeddings for your RAG app