text-embedding-3-small Notes

OpenAI's most cost-effective embedding model.

• Dimensions: 1536 (default), reducible via dimensions param
• Pricing: $0.02 / 1M tokens input
• Context: 8191 tokens
• Use case: semantic search, clustering, recommendation

Trade-offs vs -3-large

• Smaller = cheaper + faster
• Quality drop is real but small for general English
• For multilingual: -large pulls ahead

Why mdfy uses -small

50,000 docs × ~1k tokens/doc × $0.02/M = $1 to fully embed a hub. That's the right operating point for free-tier users.

Reduction trick

Truncate to 512 dims for storage savings — re-normalize before similarity search. mdfy doesn't do this yet; pgvector(1536) is fine under 100k docs/user.