Vision-language model (VLM)

The loop, in one paragraph

Give the model a frame and it writes a sentence or two about it: the people, the objects, the actions, anything hazardous. Do that for frames pulled steadily from a video stream and you get a running written account of the footage. That's the whole trick. Everything else (search, alerts, daily reports) rests on the fact that the account exists as text.

Why descriptions make footage searchable

Once footage has been described in words, the descriptions can be indexed, and finding a moment becomes a text search rather than a viewing job. The VLM writes, the index remembers, the search box reads. This is the engine behind searchable video, and it's the loop Svid is built on: plain-English queries over described frames, with results in seconds.

What a VLM doesn't need

No training on your site, no tagging, no labelled examples of your shop floor. A general model already knows what a person, a forklift and a spillage look like; you point it at footage and it starts describing. Live and recorded footage are treated identically, because to the model both are just frames. It reads images out of the stream and nothing else, so there's no audio analysis involved.