Figure 1

The same question, three deployment patterns

The useful comparison is not which label sounds safest. It is where media is processed, what crosses a boundary and how the answer returns to evidence.

Worked example — “Find the interview moment about recycled materials.”

Cloud-first

Hosted processing
Where media is processed
Provider infrastructureTranscription and analysis run in the hosted service.
What leaves the studio
Media or selected contentThe boundary depends on the service and configuration.
Where the index lives
Provider environmentLocation, retention and deletion follow the service agreement.
How the answer is verified
Only if the service returns references the studio can inspect.
Main trade-off
Rapid adoption and managed scale; the provider defines more of the data boundary.

Local-first hybrid

Foma example
Where media is processed
Close to studio storagePreviews, transcription and visual analysis can run locally.
What leaves the studio
Deployment-approved textThe question plus a limited text context; no original media.
Where the index lives
Studio-controlled environmentIndexes and locally generated derivatives remain within it.
How the answer is verified
The returned text is reattached to local evidence before it is shown.
Main trade-off
Media remains local while approved external model capability is retained.

Fully local

Separate configuration
Where media is processed
Studio-controlled environmentDefined processing components remain inside it.
What leaves the studio
Nothing in the defined flowAny network exception must be separately documented.
Where the index lives
Studio-controlled environmentIndexes and derived assets remain inside it.
How the answer is verified
The answer is linked back to locally accessible evidence.
Main trade-off
Maximum control; more studio-owned or managed capacity and monitoring.

Local-first describes an architectural priority, not a single topology. Air-gapped adds network isolation and is not a synonym for fully local. Privately deployed means single-tenant: provisioned for one studio rather than shared.