What should I use for a cheap embeddings pipeline with acceptable recall?

If you want cheap embeddings with “good enough” recall, the best choice is usually:

Recommended default

sentence-transformers/all-MiniLM-L6-v2

• Very cheap / fast
• Small model (~22M params)
• Works well for general semantic search
• Common baseline for production prototypes and budget systems

Why it’s a good fit

• High throughput on CPU
• Low memory use
• Strong quality for its size
• Easy to deploy locally or on a modest API budget

---

If you want a bit better recall

BAAI/bge-small-en-v1.5

• Often better retrieval quality than MiniLM
• Still relatively cheap
• Good choice for English-only search

If you can afford slightly more compute, this is often the better “cheap but decent” option.

---

If multilingual is needed

intfloat/multilingual-e5-small

• Good for multilingual retrieval
• Still lightweight
• Use the right prompting format for E5 models, e.g.:
  • query: "query: ..."
  • document: "passage: ..."

---

Practical pipeline for cheap + acceptable recall

If cost matters, don’t overspend on embeddings alone. Use:

1. Small embedding model
   • MiniLM or BGE-small
2. ANN vector index
   • FAISS, hnswlib, or pgvector
3. Chunking
   • Use sensible chunk sizes, e.g. 200–500 tokens
4. Reranking for top results
   • Cheap embeddings + a small cross-encoder reranker on top 20–50 results can improve quality a lot

This usually gives much better recall/precision tradeoff than using a large embedding model everywhere.

---

If you want the absolute cheapest acceptable setup

• Model: all-MiniLM-L6-v2
• Index: FAISS HNSW or IVF
• Reranker: optional, only if quality matters
• Chunk size: 300–400 tokens with overlap

---

Quick rule of thumb

• Lowest cost: all-MiniLM-L6-v2
• Best cheap quality: bge-small-en-v1.5
• Multilingual cheap: multilingual-e5-small

If you tell me:

• your language(s),
• data size,
• latency budget,
• and whether this is for semantic search / RAG / recommendations,

I can suggest a more specific stack.