Frames, Not Files

Ask someone what they are looking for and they will describe a moment, not a container.

“The shot where the car crosses the bridge.” “The part where she opens the envelope.” Nobody has ever said “the file MVI_4821.MOV”.

Yet almost every media tool indexes files. You get a thumbnail, a filename, a creation date — and then you scrub. The search succeeded and you still have to go find the thing yourself.

Seconds are the unit of truth

Video is a timeline with objects on it. If the index only knows about the container, every query degrades into manual scrubbing, which is exactly the work search was supposed to remove.

Structure it at the scene level instead:

• Shot boundaries give you discrete units that match how people describe footage.
• Per-shot embeddings let a sentence land on a segment instead of a whole file.
• Timecodes turn a result into a destination. A hit you cannot seek to is a thumbnail, not an answer.

Once those exist, a result stops being “here is a file” and becomes “here is the second it happens”.

Modalities should agree

Real archives are not all visual. A documentary cut might be found by what it looks like, what is written on screen, or what someone says in it — and those three live in different indexes.

Keeping them separate is a mistake. Each one alone is a partial answer, and the user is the one stuck reconciling them. Rank across all of them, then explain which evidence won:

“Found in dialogue at 00:14:02” is a much better answer than a ranked list with no reason attached.

The test

The bar is simple and unforgiving. Describe a moment you remember badly, in words that appear nowhere in the file, and land on it in one attempt.

Anything less and you have built a browser, not a search engine.