Create a conversation: an instruction, an output schema and typed context, sent once.

Returns a new conversation_id on every call. The context is prepared once so each evaluate sends only the turn. media_binding attaches per-turn emotion evidence from one of your live realtime sessions or one of your own processed recordings; a recording must be fully processed (409 media_not_ready otherwise), and one with several subjects needs a selection per modality. Alternatively, a recording binding may name one to four people, each with a unique role label (1–64 trimmed characters) and explicit subjects. Each person's readings stay separate. people and binding-level subjects are mutually exclusive; overlapping subject IDs within a modality are 422. In people mode omitted modalities remain unbound, even with a sole speaker. This also works on each entry of media_bindings. Labels and selections persist for the conversation lifetime.

With retention: ephemeral no turn's input or output is stored, only its usage. The conversation's own instruction, schema and sections are kept until it expires or is deleted, so put per-turn content in evaluate, not here.

Also available on a user's delegated access token with the insights.write scope. The conversation then belongs to that user and your app: only your app's tokens for that user can evaluate or delete it. It must be bound to one of that user's own processed recordings, which also needs media.read (403 naming the scope otherwise); a conversation with no binding, or bound to a live session, is refused with 403. Each turn is billed to that user, or to their team workspace where the deployment is that workspace's own.

A delegated token may instead send media_bindings: one to 200 of that user's processed recordings, each with its own media_id (distinct ids; more than 200, a repeat, or sending media_binding as well is 422). The conversation then reads several recordings at once, and each turn chooses which with media_ids. A recording is readable by that user as its owner: their own, or those of the team workspace the token's grant is bound to. Every id is checked before anything is read or charged, and one that is missing or not readable, among ones that are, fails the whole request with 404 binding_not_found; then media.read is needed (403 naming the scope), and every recording must be fully processed (409 media_not_ready, with each unfinished media_id and its status). media_bindings is refused with 403 on an API key. Turns on recordings of the user's own are billed to that user. A grant bound to a team workspace is billed to that workspace where the workspace has its own dedicated deployment, and is refused with 403 anywhere else. The whole selected cohort must become ready within one readiness budget; incomplete preparation answers 409 media_not_ready without creating a conversation. If recording preparation is temporarily unavailable, the request answers 503 preparation_busy; retry after a short wait.