Core concepts

A JSON Schema describing the typed output you want from a video. You register a schema once with POST /v1/schemas and reference it by id from any job.

One extraction run on one source video. A job pairs a source_uri with a schema_id, moves through a queued and running lifecycle, and produces a result.

The typed JSON Lirovo extracted, conforming to the schema you supplied. Fetch it with GET /v1/jobs/{id}/result once the job is done.

Every extracted value points back to a source moment, carrying a timestamp and a modality (audio or visual). Evidence is not optional: it is how you verify and cite each field.

During extraction, Pass A builds a compact knowledge graph of nodes and edges with evidence-linked spans. The graph is mirrored into the metadata store and is callable for traversal alongside the result.

The intermediate products of the pipeline: the transcript, the frame inventory, and the per-frame vision descriptions. Each is retrievable on its own, so you can inspect exactly what the models saw and heard.

Where results are pushed when you want delivery rather than polling. The first destination kind is a signed webhook (HMAC-signed, retried with backoff).

Lirovo is multi-tenant from day one. Every schema, job, result, artifact, and destination is scoped to your tenant; there are no cross-tenant reads. Your API key resolves to the tenant that owns the resources it can touch.