Context window
Also known as: context length, input window
The amount of material a model can consider at once, which decides whether a long upload is judged as a whole or in separate pieces.
Updated
What it means
A context window is the ceiling on how much a model can hold in view for one answer. For video analysis it decides something concrete and visible in the result.
Either an hour of footage can be reasoned about together, or it has to be chopped into chunks that are each judged in ignorance of the others. The second case is far more common than people assume.
The difference shows up in what gets chosen. A model that sees a whole session can notice that the fourth attempt was the one that worked. A model shown ninety seconds at a time cannot, because the earlier attempts were never in the same view.
How it works in practice
Long material is usually handled in overlapping passes. Each chunk is analysed, its findings are summarised, and the summaries are compared at the end, which keeps the window from being a hard limit on upload length.
Frame sampling does the other half of the job by reducing how much has to fit in the first place, since a sampled clip occupies a small fraction of what the full recording would.
What goes into the window is a real choice. Spending it on more frames from fewer clips, or fewer frames from all of them, produces noticeably different edits from identical footage.
What to watch out for
Chunking quietly breaks comparisons. Picking the best wave of a session is impossible if no pass ever saw more than two of them, so the system returns the best wave in each chunk instead.
Nobody notices the difference until the edit repeats itself, at which point it looks like a failure of Duplicate shot detection rather than a consequence of how the material was divided up.
Filling the window is also not free. More input means a slower answer, which pushes up Model latency on every job, and beyond a point the extra material stops improving the choices at all.
How RawClip handles it
Long uploads are analysed in overlapping passes whose findings are compared at the end, so the strongest moment of a session is chosen against the whole session rather than against whatever happened to sit in the same chunk of it.