Copilot
The Copilot answers questions about activity at your site in plain English. Open the Copilot panel on the right of the dashboard, type a question and press Enter. Anyone who can sign in can use it.
The Copilot runs on your site node, with a local language model served by Ollama. It reads your site's events, alerts, cameras, detection profiles, workflows and edge node status, and nothing else.
How it answers
The Copilot never gets direct access to the database. It answers through a fixed set of read-only tools:
| Tool | What it looks up |
|---|---|
| Count alerts | How many alerts match a status and, optionally, a camera. |
| List alerts | Recent alerts, by status and camera. |
| Count events | How many detections or rule events match a camera, object class and time window. |
| Query events | Recent detections or rule events, with the same filters. |
| Search by description | Alert clips whose captions are closest in meaning to your words. |
| Get event | The details, caption and clip link of one event. |
| List cameras | The cameras, their detection profiles and their latest stream health. |
| List detection profiles | Each profile's classes, trigger, schedule and severity. |
| List workflows | Each workflow's condition, steps and whether it is enabled. |
| List edge nodes | The edge nodes and whether their heartbeat is current. |
| System health | A summary of cameras, edge nodes, open alerts and event totals. |
Common questions, such as alert counts, detection counts, system status and the camera list, are answered straight from a database query without waiting for the model. Other questions go to the model, which picks a tool and phrases the answer.
When an answer mentions an event that has a clip, it includes a link such as /clips/134299. Click it to watch the clip. The Copilot can't play or attach video itself.
Example questions
These are answered straight from the database:
- "How many open alerts are there right now?"
- "How many resolved alerts are there?"
- "List the most recent alerts"
- "How many people detections in the last hour?"
- "How many bags/backpacks have been seen?"
- "Get me the details for event 1234"
- "How many cameras are there?"
- "What is the system status and health?"
The panel also has quick prompts for status, alerts, cameras, recent events, vehicles, rules and the dashboard.
How it reads your question
- Alert status. Alert questions are about open alerts unless you say "acknowledged", "resolved", "dismissed", "closed" (resolved and dismissed) or "all". "List the most recent alerts" lists open alerts. "List all recent alerts" lists every status.
- Object classes. The detector stores precise class names, such as
personorhandbag. "People" and "person" meanperson. "Bag", "bags", "backpack" and "backpacks" meanbackpack,handbagandsuitcase. - Time. It understands phrases such as "in the last hour", "in the last 3 hours" and "today". "Today" means the last 24 hours.
- Cameras. Name cameras as they appear in the Cameras & Edge Nodes tab. The Copilot matches whole words of your camera names, so "loading bay" finds a camera called
loading-bay.
A camera named after a common word, such as gate or door, is picked up whenever that word appears in a question. Prefer distinctive camera names, such as north-gate or cam-07.
Search by description
To find something by what it looks like, ask a "Did anyone..." question with visible details, for example:
- "Did anyone wear a red jacket recently?"
- "Did anyone have a tripod or camera equipment recently?"
The Copilot compares your words with the captions of alert clips and returns the closest matches with their event IDs and clip links. The model decides when to search by description, so phrasing matters. "Did anyone..." works better than "Was there a...?".
Search by description only finds alerts that already have a caption. It can't find detections that never raised an alert, because Argus only captions alert clips. See Video captions.
Limits
- Short follow-ups. The panel sends the earlier messages of your conversation with each question, so follow-ups such as "and the resolved ones?" often work. For precise answers, put everything the Copilot needs in the question, such as the camera name or event ID. Reloading the page starts a new conversation.
- Read-only. It can't acknowledge alerts, change rules or add cameras.
- No filtering by workflow name. The Copilot doesn't count or list alerts by workflow name. Ask by camera or status instead.
- Short answers. It looks at a limited number of results for each lookup. For long lists, use the Alerts tab or the API.
- Model limits. Questions that go to the model can get a less precise lookup, especially yes-or-no questions. If an answer looks wrong, rephrase it and check the Alerts tab.
If the model is busy or slow, the Copilot answers with a short summary of the latest open alerts and events instead of a full answer. The camera line in that summary is not a live count. Check your cameras in the Cameras tab.
Models
The Copilot uses a local language model served by Ollama on your site node. Questions and answers never go to a cloud service. The model is configurable: the default is gemma3:4b, and your administrator can set another Ollama model on the site-api service with COPILOT_MODEL. See Install. For help choosing a model for your hardware, contact us.
Copilot questions and answers are not stored on the site.
Use it from your own tools
Send questions over HTTP with POST /copilot/chat.