Multipart upload. Repeat with more files — every batch_size-th chunk auto-creates a batch.
No request sent yet.
Paginated, ordered by batch_index. Click a row to load it into the next step.
No request sent yet.
404 TRANSCRIPTION_BATCH_NOT_FOUND also fires if the batch doesn't belong to this session.
No request sent yet.
The audit trail behind a status — merge, submit, and cleanup events, in order.
No request sent yet.
Available once the batch reaches TRANSCRIBED. 404 TRANSCRIPT_NOT_FOUND otherwise.
No request sent yet.
Generated automatically once a batch reaches TRANSCRIBED — cumulative
per session, not per batch. 404 CLINICAL_NOTE_NOT_FOUND if MedGemma
isn't configured yet or hasn't finished. Citations show both an exact-match
verified flag and a fuzzy match_score (small models rarely
quote byte-for-byte) — medical_history/recommendations/
llm_contraindication_warnings/missing_questions are the
model's own inference, not transcript-grounded.
No request sent yet.
Zero-LLM, keyword-rule engine — recomputed from the full transcript-so-far every
time a batch transcribes. Deliberately separate from the clinical note's own
llm_contraindication_warnings above; an empty list here is a normal,
valid result, not an error.
No request sent yet.
Stops new chunks from being accepted. Already-uploaded chunks keep processing.
No request sent yet.
Mints a bearer-token link a phone can scan to become a remote mic for this
session — recording and uploading go through the exact same pipeline as step 2,
just token-authenticated instead of device-key. No consent gate, no expiry
(matches the reference implementation this was ported from). The QR itself is
rendered by GET /qr?data=..., a stateless local PNG encoder — no
external QR service.
No request sent yet.
Request log
Every call this page has made this session, newest first. request_id is what to grep in the server logs.
- Nothing sent yet.