What should I use for embeddings if I have frequent re-indexing?

If you expect frequent re-indexing, the best embedding choice is usually:

Use a fast, cheap embedding model

because the bottleneck becomes embedding generation cost and latency, not retrieval quality alone.

Good default options

• OpenAI text-embedding-3-small
  • Fast
  • Low cost
  • Good enough for many search/RAG use cases
• If you need higher quality and can afford more cost:
  • text-embedding-3-large
  • Better semantic quality, but slower/more expensive

Why this matters

If you re-index often, you’ll repeatedly pay for:

• embedding every document/chunk
• storing/updating vectors
• re-running similarity search indexes

So the tradeoff shifts toward:

• lower embedding cost
• shorter indexing time
• smaller vectors

Practical recommendation

• High-frequency re-indexing + large corpus: text-embedding-3-small
• Less frequent re-indexing or quality-critical retrieval: text-embedding-3-large

Extra tips

• Keep chunks reasonably sized to reduce total embeddings.
• If only part of your data changes, do incremental re-indexing instead of rebuilding everything.
• Use a vector DB/index that supports upserts efficiently.

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

• corpus size
• update frequency
• latency/cost constraints
• whether this is for RAG, search, or clustering