Where it is: Settings → General & AI → Embeddings.
Two different models are involved in every answer. The chat model writes the reply. An embedding model turns your content and each question into vectors so the right passage can be found. They can come from different providers.
Your options
- OpenAI — the default, and the safe choice.
- Voyage AI — a generous free tier, useful if your chat model is from another provider and you would rather not open an OpenAI account.
Cost
Embedding is cheap and mostly one-off: your content is embedded once, and unchanged content is never re-embedded. Ongoing cost is one small embedding per question, and repeated questions reuse a cached vector.
Changing the embedding model later means the existing vectors no longer match. System Health warns you, and Re-embed all in Settings rebuilds them. It is the single most expensive button in the plugin, so change the embedding model rarely.