Lloyaldocs
Video

Configuring video

Every setting has bounds, and a value outside them is refused. Here is what each one trades, stage by stage.

On this page

Each key's bounds, default and when a change applies are on harness.yml, read off the tables the keys are declared in. This page is what each setting does to a video.

Taking a clip

Setting What it trades
attachments.video.fps How often a frame is sampled, up to the frame cap. More frames see more of a short clip and cost more cells.
attachments.video.maxFrames The most frames a video keeps. A clip longer than the cap allows at fps is sampled evenly instead, so the cap decides how much of a long clip is seen.
attachments.video.maxEdgePx The longest edge a frame is scaled to: detail against what each frame costs.
attachments.video.soundtrackWindowSeconds How long each window of the soundtrack runs. Each is transcribed on its own and becomes one caption, one span of search by sound, and one note beside the frames. Shorter is finer; longer means fewer transcriptions. It doesn't follow the frames' time marks.
attachments.video.maxDurationSeconds The longest clip a harness takes, refused before any byte is pushed. Unset, a clip of any length is taken; length costs upload and one transcription per soundtrack window.

These are set in harness.yml and fixed for the run: the page samples a clip by the policy the host described to it, so a change takes a restart. The time marks between frames are not a setting: they follow the frames' spacing, so frames keep fusing at any length.

Hearing it

Setting What it trades
model.transcription Its presence transcribes each soundtrack window when a video is composed. Absent, a video arrives with no transcript.
model.transcription.maxBytes, maxDurationSeconds, maxSampleRate The limits a soundtrack window is cut and written to meet, so every window is one the service admits.

When an embedding model that hears is bound as well, the page cuts each window to meet what both admit.

Seeing it

Setting What it trades
model.vision.minTokens, model.vision.maxTokens What one frame costs the reasoning model. A frame is an image, so this is the main lever on what a video costs.
attachments.context The share of the context a turn's attachments may fill. A video over it is refused before anything is read in.

The harness also passes admitted a reserve, the room a run needs once the media is in, and free, the room a new conversation will have. Both are the harness's own, from the budgets its runs use. Cap what attachments take is how.

Embedding it

Setting What it trades
model.embedding Which model places a video in one space with text, images and sound; EmbeddingGemma 2 when it names none.
model.embedding.context The most cells one input may take, and so the longest window that can be embedded whole. A longer input is refused, never truncated.
model.embedding.vision.minTokens, maxTokens What one frame costs the embedding context: detail against how many frames fit in one input.

The host and the page

Setting What it trades
The content plane's maxUploadBytes, 8 MiB by default The largest single attachment posted whole. A video's original is pushed in chunks instead.
The blob surface's chunk ceiling, 64 MiB The largest chunk of an original one request may carry.
BlobUploads(…, { registry: true }) Off by default: the host admits only a video's original. On, it is a full registry that admits and serves any blob.
The composer's accepted types Which files a reader can drop: add video/*.
The desktop Content-Security-Policy media-src attachment: blob:, so the player can play the original and its tracks.
↑
Search