Wiring & DAG

AVFlow builds a directed acyclic graph (DAG) from each component’s inputs array. Each entry may be an upstream component name or an object with name and an edge-specific select filter.

Basic rules

1. Sources have no inputs — they produce media.
2. Nodes consume one or more upstream components and produce processed media.
3. Sinks consume upstream components and write output (stream, file, room, etc.).
4. Every inputs entry must reference an existing component by name.
5. The graph must be acyclic.

Media types

Components exchange video, audio, and optionally captions (from ASR). AVFlow infers which streams to pass when you connect components — you rarely need manual split nodes.

Use select to filter by participant identity, track type, or media type. On RTC sources it lives inside config; on nodes and sinks it lives on the input entry.

Use object-form inputs to filter each upstream:

"inputs": [
  { "name": "room_a", "select": { "identities": ["alice"] } },
  { "name": "room_b", "select": { "identities": ["bob"] } }
]

A filter applies only to its own edge. What passes it is still limited to the media types the component consumes.

Common patterns

Source → sink (relay)

A relay has no processing nodes: the source connects straight to the sink. This works for pull sources (rtmp_pull, whep, srt_pull, rtsp_pull, file_pull) and 1:1 sinks such as rtmp_push.

rtmp_pull ──► rtmp_push

Source → mixers → sink (mix)

RTC room sources (livekit, jitsi, daily, agora) are multi-stream (1:n) and need a mix — video_mixer and audio_mixer — before rtmp_push or similar 1:1 sinks, even for a single participant.

Composite recording

Same mix pattern as above, often with a segment sink:

livekit ──► video_mixer ──► rtmp_push
        └──► audio_mixer ──┘

ASR captions

Wire the ASR node into sinks that support subtitles (segment, rtmp_push, RTC sinks, etc.).

livekit ──► audio_mixer ──► segment
        └──► asr ──────────┘

Per-track audio over WebSocket

Both audio_resample and websocket are n:n, so RTC participant tracks stay independent:

livekit ──► audio_resample ──► websocket
  1:n              n:n              n:n

Validation errors

Jobs are rejected at submit time if:

• A 1:1 RTC sink (e.g. rtmp_push) is wired directly to a multi-participant RTC source without a mixer
• segment is connected directly to a multi-track RTC source (use mixers first)
• Input limits are exceeded

Fix the topology and resubmit — AVFlow does not partially run invalid graphs.

Changing wiring on a running job

Resubmit the full job with the same name. You can add or remove sources, nodes, and sinks, and update inputs on existing nodes and sinks in the same request — for example, add a second livekit source and append it to a video_mixer’s inputs array.

Sink input changes use make-before-break switching: new upstream tracks are connected before old ones are removed, so output stays continuous when possible.