what should I use for embeddings if I need frequent updates?

If you need frequent updates, the best embedding choice is usually one that is cheap, fast, and stable enough to recompute often.

Good options

• Open-source sentence embedding models (best if you want control and low cost):
  • sentence-transformers models like all-MiniLM-L6-v2 for speed
  • bge-small / bge-base
  • e5-small / e5-base
• Smaller embedding models are better than large ones when updates are frequent, because you’ll regenerate embeddings often.
• If you’re using a hosted API, pick a lower-cost embedding model rather than the most capable one, unless quality is critical.

What matters most for frequent updates

1. Low latency / low cost per embedding
2. Easy batch processing
3. Consistent embeddings over time
4. Good enough semantic quality
5. Incremental re-embedding support in your pipeline

Practical recommendation

• For most use cases: bge-small or all-MiniLM-L6-v2
• If you need better quality but still want frequent refreshes: bge-base or e5-base
• If you need very high volume updates: consider a self-hosted embedding model so you’re not paying per request

Rule of thumb

• Frequent small updates → use a small, fast embedding model
• Rare updates but high retrieval quality needed → use a stronger model
• Constantly changing data → design your system to re-embed only changed chunks, not everything

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

• cloud vs self-hosted
• document size
• update frequency
• budget
• vector DB you’re using