what embedding model should I use for document search?

For document search, the best embedding model depends on your constraints, but a very strong default today is:

Good general-purpose choice

• OpenAI text-embedding-3-large — best quality for semantic search in many cases.
• If you want lower cost / faster:
  • OpenAI text-embedding-3-small — solid quality and much cheaper.

Good open-source alternatives

• BAAI bge-large-en-v1.5 — excellent for English retrieval.
• intfloat e5-large-v2 — also very strong, widely used.
• multilingual:
  • bge-m3 — strong multilingual and retrieval-capable choice.

Practical recommendation

• If you want the best out-of-the-box quality and don’t mind API usage: text-embedding-3-large
• If you care about cost/performance: text-embedding-3-small
• If you need to self-host: bge-large-en-v1.5 for English or bge-m3 for multilingual

Important note

For document search, the embedding model alone doesn’t determine performance. You’ll usually get a much better system if you also:

• chunk documents well
• use hybrid search (BM25 + embeddings)
• rerank top results with a reranker model

If you want, I can recommend a model based on your setup:

• OpenAI vs open-source
• English-only vs multilingual
• small/cheap vs best quality
• RAG on PDFs, docs, code, or web pages