The problem
We came back from Japan with a terabyte of footage across 4,584 clips, spread over
nine shoot locations. Filenames were IMG_2532.MOV. Folder names were dates.
Finding the shot you remember in that is not a search problem, it is an archaeology problem, and it silently kills a content pipeline: the footage exists, and the cost of locating it exceeds the value of using it.
What it does
I describe a shot the way I would describe it to a person, and it hands me the clip. “Morning light on a tea field” returns the Kakegawa pans.
How it works
Every clip is sampled, described by a vision model, and embedded alongside a CLIP embedding of the frame itself. Both go into a local vector store. A query is embedded the same way and matched against the library.
The part worth stating plainly is the cost.
An index that costs six dollars is one you re-run whenever the model improves. An index that costs six hundred is one you build once, regret, and never touch again.