Ask about the Conversational AI course material. Answers come
only from the evidence packs built from prof/ — the question is
routed to the packs it needs with BM25, and those packs are cached, so a
repeat topic reads from cache instead of re-sending the store. Follow-up
questions carry a summarised digest of earlier turns, so
“and for 10 million?” resolves against what you just asked.
Conversation
Both on is the measured default: ~$0.011 and ~1.1s to
first token. Turn both off to send the whole store with the question in
the system prompt — ~$0.386 and ~3.5s.