It varies by feature — several are free, and the paid ones are metered so you can see the spend before committing.
Free at run time: structure-aware chunking, Hybrid Search, and Entity Safety. They use your database, not your AI provider.
One AI call per chunk, when you index: content enrichment (blurbs, coreference rewriting, entities, and facts share a single combined call per chunk), plus Summaries and Questions — which are on by default and each cost the same way. The Estimated indexing calls readout on a source's Retrieval Accuracy tab shows the exact count for your switch positions before you reindex.
One low-cost AI call per visitor message: the Query Planner and the reranker (either strategy). These run on the dedicated Accuracy LLM Profile, so point it at a small, cheap model — answer quality still comes from the chat's main profile. The chat's Expected accuracy calls readout totals these.
Per answer: Citation Integrity rides the normal answer request at no extra call; Strict Answer Verification adds one support check and, when a claim fails, at most one regeneration.
Diagnostics keeps a running count of accuracy calls and their average latency. For which features are worth their cost first, see Improving Answer Accuracy.
