Yes, in two specific moments — and it stays on your server the rest of the time.
- When you index content, each chunk of text is sent to your embeddings provider (OpenAI) to be converted into a vector. The vectors come back and are stored in your own database.
- When a visitor asks a question, the question plus the retrieved content chunks are sent to your chat model's provider (OpenAI or Anthropic) to write the answer.
Both happen under your own API account and that provider's data terms. Content you never index is never sent anywhere, so the practical rule is: only create embeddings from content you are comfortable passing through your AI provider, and use namespaces to keep a chat scoped to exactly that content.
What Wyregraf keeps on your own server about the conversations themselves is covered in Are visitor conversations saved?
