Ask Dardaris a question about the portfolio

Every question goes to the assistant (status: generated), which decides how to answer it: it works figures out exactly with the answering engine over the records and shows them as cards, reads the filing and the help pages, reads the portfolio through the MCP actions, and proposes any change as actions rather than making it. Where no model can answer, the engine answers from the records on its own, and a question asking for a document is put to the filing, whose match is the passage the document actually contains. When nothing can be said, answer is null and followUps becomes a handful of questions that are known to have answers.

The turn is read against the last one: send it back as previous and a follow-up carrying only a qualifier keeps the topic it inherits, so "and per building?" is read against the answer above it.

Nothing is stored unless you say where to store it. Name a conversationId and the turn is appended to that transcript, which is what the front desk does so a chat can be read again; leave it out and the request is stateless. Name a turnId as well and asking the same turn again, after an answer that was lost on the way back, writes it once.