2026-10-11

Remix tile accepts a reference ingest id; other verbs do not

Use a finished reference ingest id as understanding_id for remix tile, including around without a transcript job. Boundaries and cut still need a probe.

4 min readSume

All posts

Yes: the remix media tile verb accepts the id of a completed reference ingest as its understanding_id, and it is the only verb that does. An around phrase then works without a transcript_job_id, because the ingest manifest already carries its own words.

Every other verb, including boundaries, answers 404 for an ingest id and wants a probe. This is read from the route code and its tests on origin/main, with the two docs pages Remix media and Reference ingest, read 2026-10-11.

Why this shortcut exists

A reference ingest reads one clip of up to 300 seconds and returns shots, text tracks, audio facts and a labeled strip in one call. If the read transcribed, the manifest has the spoken words with times. Instead of running a separate probe and transcribe just to look at frames, you can point tile at the ingest you already paid for.

The route comment calls this the reference-remix path: it grids the clip it ingested, labeled with that ingest's transcript.

The rules, verb by verb

The last row is the cautious reading: the route test checks boundaries and tile, and the code opts in to ingest ids only when the verb is tile. Treat every other verb as needing a probe.

Which id each remix media verb accepts, from the route code and its test (read 2026-10-11)
Request	With a probe id	With a reference ingest id
tile with at, ranges or start/end	Accepted once the probe is complete	Accepted once the ingest is complete
tile with around, no transcript_job_id	400 remix_tile_transcript_required	Accepted
tile before the job completes	404, not ready yet	404, not ready yet
boundaries	Accepted	404
transcribe, cut, notes, align	Accepted	Not accepted for ingest ids

A short sequence

With a completed ingest, the sequence is one read and then as many tiles as you need.

Run POST /v1/reference-ingest with an Idempotency-Key. The default mode waits up to 30 seconds for the manifest; otherwise it returns 202 and you poll the job.
Take the job id and send it as understanding_id to POST /v1/remix-media/tile with around: { phrase: "..." }.
For a dense look at a cut, add start, end and every, staying under 96 cells per call.
If you need a cut candidate series or a labeled excerpt, run probe on the same clip and use that id for boundaries and cut.

Watch the speech setting

around on an ingest id relies on the ingest's words. An ingest that did not transcribe has none: speech.allow_billed_stt defaults to true for purpose: reference_remix and false for the other purposes, and the speech step is skipped when the track is silent or has no speech. If the ingest ran with STT off, run transcribe on a probe and tile from that instead.

The ingest bills its Modal compute and, when it transcribes, the speech-to-text per-minute rate on the hint, then settles to what ran, as described in the Reference ingest pricing section.

When to run a probe anyway

The shortcut saves a probe and a transcribe when all you need is pictures with words on them. It does not give you a boundary series, a labeled excerpt from cut, a durable 16 kHz wav or the notes verbs. If your plan includes any of those, start with probe and run transcribe in the same batch, as the call-order post describes.

A second reason is timing: a tile on an ingest id waits for the ingest to finish. If the read returned 202, poll it first, because tiling a queued ingest answers 404 and not a queue position.

Where this runs

Both surfaces are flag-gated. The docs say remix media follows SUME_COM_REMIX_MEDIA_ENABLED and reference ingest follows SUME_COM_REFERENCE_INGEST_ENABLED; development has each on and production is opt-in. The ingest-id shortcut needs both flags on in the same environment.