audio-dispatch

Audio dispatch

voxtral_transcribe and mistral_chat are available in core, metier-docs, and admin. Check mistral://capabilities when a tool is missing.

Transcribe

Use a provider-accessible audio URL:

{
  "audio": { "type": "file_url", "fileUrl": "https://example.com/meeting.mp3" },
  "diarize": true,
  "timestampGranularities": ["segment"]
}

For an existing audio upload, replace audio with { "type": "file", "fileId": "<audio file ID>" }. Add language: "fr" or "en" only when known; otherwise omit it. A local path is not a URL. Obtain an accessible URL, an existing audio file ID, or a transcript if no upload route is available. This server's files_upload schema does not accept purpose: "audio".

Check isError first. Read structuredContent.text and optional segments[]: each segment has text, start, end, optional score, and optional speaker_id. Preserve timestamps and speaker IDs in the transcript. Do not map IDs to people without evidence; if diarization is absent, report that attribution is unavailable.

Extract decisions and actions

Pass the labeled transcript to mistral_chat as string content in messages. Omit model to use the configured default. Ask it to extract explicit commitments, decisions, and unresolved questions, preserving speaker IDs and timestamps. Mark unspecified owners and deadlines as unknown, and distinguish proposals from accepted decisions.

For JSON output, use response_format: {type: "json_schema", json_schema: {name: "meeting_dispatch", schema: <JSON Schema object>, strict: true}} with a schema for the requested fields. Parse the JSON from structuredContent.text; it is not a separate typed dispatch payload.

Present actions by speaker with any stated deadlines, followed by decisions and open questions. Check attribution against the transcript. The french-meeting-minutes skill can format a requested French report from the same transcript. Custom intent labels can be requested through chat; there is no need to assume a classifier or batch deployment.