Chunking

Four boundary signals fused into cut points — the single biggest lever on how good the analysis turns out.

A chunk is the span every analyzer runs on and every vector describes. Cut in the wrong places and a scene description straddles two unrelated moments, a sentence is split mid-clause, and retrieval returns spans that are technically correct and practically useless.

So chunking is asked about at upload time rather than defaulted silently.

The four signals

Each detector runs once over the whole video and emits (time, strength) events.

Signal	Detector	Fires on
`speaker`	pyannote	A change of speaker
`silence`	Silero VAD	A gap in speech
`cut`	PySceneDetect	A hard visual cut
`semantic`	CLIP frame embeddings	Visual drift — the shot is now of something else

Not every video offers every signal. Unbroken CCTV has no cuts; silent footage has no speaker turns; a static lecture recording has neither.

Fusion

Events are combined into a single score curve over the video: each contributes a Gaussian bump of height weight × strength, so boundaries several detectors agree on reinforce each other. Peaks are then picked greedily, strongest first, with a minimum gap enforced so the result is not a run of back-to-back fragments.

Weights renormalise over the signals that actually fired. Without that, a preset doubles as a granularity dial: on single-signal footage, audio and video produced 12 versus 63 boundaries from identical content. Renormalising means a preset expresses which signals matter, not how many cuts to make.

Three modes

A named weighting of the four signals. Each preset's weights sum to 1.0.

Preset	speaker	silence	cut	semantic	Best for
`audio`	0.35	0.35	0.15	0.15	Podcasts, calls, lectures — anything carried by speech
`video`	0.15	0.15	0.35	0.35	Surveillance, silent footage, b-roll, sports
`audio_video`	0.25	0.25	0.25	0.25	Most footage — interviews, meetings, edited video

{ "mode": "preset", "preset": "audio_video", "min_duration": 5, "max_duration": 20 }

Duration bounds

min_duration (default 5s) and max_duration (default 20s) clamp the spans produced by fusion. They apply to preset and weights, never to interval.

Short chunks give precise timecodes but thin context per vector; long chunks give richer descriptions but vaguer citations. The defaults sit where a scene description still reads as one coherent moment.

Choosing in the app

The upload dialog exposes all three modes, with the preset options labelled by what they suit. The chosen config is stored on the row as ingest_config and replayed verbatim on re-index, so "re-run this the way it was run before" needs no extra state.

A one-line summary appears on status surfaces:

Audio + video · 5–20s · 2 analyzers
Fixed 20s chunks · 3 analyzers
Custom weights (speaker, cut) · 5–20s · 2 analyzers

Changing it later

Re-indexing with a different chunking replaces the chunks and drops every vector that described the old ones — they no longer describe anything that exists. Re-indexing with the same chunking keeps what already ran and adds any newly requested analyzers.

Because chunk_config is part of a vector's identity, one video can hold several chunkings side by side, and searches can be filtered to one of them with the chunk_config filter.