Skip to main content

Video captions

For each alert clip, Argus writes a one-sentence description of what is visible, for example "A person wearing a gray shirt is walking past the entrance." You see the caption on the alert card, and the Copilot uses it to find clips by description.

How captions are made​

  1. A rule matches and the edge node saves the clip.
  2. The captioning service takes the middle frame of the clip, the moment the rule matched.
  3. It sends the frame to a vision-language model, with the camera name, workflow name, detection profile and object class as context.
  4. The model returns one factual sentence. Argus saves it with the event.
  5. Argus also turns the caption into a text embedding and saves it. This is what makes search by description work.

The model runs in Ollama on your site node. The frame never leaves the site.

Captions are written only for alert clips, never for the live stream. The caption describes the alert. It never decides whether an alert is raised. The detection profile and rules decide that.

Caption status​

StatusWhat the alert card showsMeaning
PENDINGCaption pending...The clip isn't ready, or captioning hasn't started.
PROCESSINGGenerating description...The model is working on it.
COMPLETEDThe captionDone.
FAILEDCaption unavailableThe clip couldn't be read or the model didn't answer in time. The alert itself is not affected.

Captioning takes one clip at a time, in the order the clips arrive. When many alerts fire at once, captions can lag behind the alerts.

What captions say​

The model is instructed to:

  • Write exactly one short sentence.
  • Describe only what is visible: colours, such as "red car" or "white van", and actions, such as "is entering" or "is standing near".
  • Use neutral wording, such as "a person wearing a gray shirt".
  • Never guess age, gender, identity, intent, emotions or criminality.
note

Captions come from an AI model and can be wrong or miss details. Use them to find and sort clips, and watch the clip before you act on an alert.

Models​

Captions and search embeddings come from local models served by Ollama on your site node. The models are configurable: by default Argus uses gemma3:4b for captions and nomic-embed-text for the search embeddings, and your administrator can point the captioning service at another vision model in Ollama. See Install.

If the embedding model isn't available, captions are still saved, but those clips can't be found by description.

Was this page helpful?
On this page
Talk to our teamQuestions about licensing or sizing VoisX Argus for your site? We can help.Book a site assessment →