Wyregraf itself doesn't charge per use — the costs are your AI provider's API fees, and they come in two kinds:
- Indexing (one-time per index run): each chunk of content is embedded once. Embedding models are cheap — indexing even a large site typically costs cents. The embedding's Index tab shows the token count and an estimated cost before you run it.
- Chatting (ongoing): each visitor question makes one chat-model call, and its cost scales with the model you picked in the LLM profile and how much context the chat hands it (the Top K setting). This is the cost worth watching on a busy public page.
Three levers keep chat costs predictable: choose a mini-tier model (retrieval assistants rarely need a flagship), bound each exchange with the Message Limit and History Limit on the chat's Restrictions tab, and/or restrict the chat to logged-in users. Enabling Summaries & Questions adds LLM calls at indexing time, so expect those index runs to cost more than standard ones.
