Semantic video search
Also known as: natural language video search, content based search
Finding footage by describing what is in it, rather than by filename, folder or the date the camera happened to record it.
Updated
What it means
Semantic video search answers a request phrased in ordinary words. Asked for the clips where somebody falls in shallow water, it returns those clips without anybody having tagged them first.
It works by turning both the footage and the request into numbers in the same space, so closeness between them means similarity of meaning. That representation is what Video embeddings provide.
The practical appeal is obvious to anybody with years of cards. Filenames carry a date and nothing else, so without search the only index is memory, and memory of a season of riding is not a usable index at all.
How it works in practice
Indexing happens once, at ingest. Each clip is reduced to a vector, stored, and compared later by Vector similarity, so searching is fast even when building the index was slow.
Granularity decides usefulness. Indexing whole files finds the right file and leaves somebody scrubbing through forty minutes of it; indexing stretches finds the actual moment, which is the difference between a search that helps and one that merely narrows the folder.
Combining it with cheap filters works better than relying on it alone. Narrowing by date, activity or duration before the semantic step keeps results relevant, and Activity classification supplies exactly that kind of coarse filter.
What to watch out for
Results are ranked rather than correct. There is no boundary between matching and not matching, only an ordering, so a search always returns something whether or not the footage exists.
Descriptions of absence fail outright. Asking for clips without people tends to return clips with people, because the representation captures what is present rather than what is missing.
Vocabulary mismatch is the quiet failure. Footage described internally as a jump will not surface for a request about a drop, which is the same brittleness that affects a Prompt and is just as hard to notice from the outside.
How RawClip handles it
Footage is indexed by what it contains as it is analysed, which is how the system locates comparable moments across a long upload instead of treating each file as an unlabelled block of time. The same index is what lets repeated attempts at one trick be recognised as repeats.