What should I use for embeddings if I care most about recall?

If you care most about recall, the best choice is usually:

1. A strong, general-purpose embedding model
2. A vector index/search setup tuned for high recall
3. Often a reranker on top, if latency allows

In practice

For embeddings alone, choose a model that is:

• high quality / state of the art
• trained for semantic similarity and retrieval
• large enough to capture nuance

Examples of commonly strong choices:

• OpenAI text-embedding-3-large for very strong retrieval quality
• bge-large / bge-m3
• E5-large / multilingual E5
• GTE-large

Important point

If you mean recall in search results, embedding choice is only part of it. You’ll usually get better recall from:

• higher-dimensional or larger models
• searching with larger efSearch / numCandidates / nprobe / similar parameters
• using hybrid retrieval (BM25 + vectors)
• reranking after retrieving more candidates

Rule of thumb

• Best raw embedding quality for recall: use one of the top-tier large models
• Best system recall overall: embeddings + generous ANN search settings + hybrid retrieval + reranker

If you want, I can recommend a specific embedding model depending on whether you’re doing:

• semantic search
• RAG over documents
• code search
• multilingual retrieval
• on-prem/open-source only