Assistant
The assistant is the console's conversational surface: a composer at the bottom of the scene that answers operational questions by driving the same deterministic engines and UI the rest of the console uses. It can select entities, seek the timeline, run analyses, build charts, import files, and synthesize scenarios, and everything it does is visible as it happens.
This page covers the chat surface itself. The full capability catalog (every tool, query patterns, the response contract) lives in the Agentic section, with worked examples in Agentic examples.
The chat stack over the globe: a question in the composer, a streamed answer above it with an expanded tool-call trace, and an insight card stating scope, window, and assumptions.
Context: pins and mentions
The assistant does not guess what you are looking at; context is explicit and inspectable.
- Context pins. Selecting a node or link on the globe automatically pins it to the conversation, so "why is this link degraded" refers to the thing you clicked. Pins are visible above the composer and removable individually.
- @-mentions. Type
@in the composer to reference any asset by name without selecting it on the globe: satellites, gateways, imported assets, catalog objects. Mentions resolve through the same registry as search. - Scene context. Every turn carries the current timeline instant, active filters, and selection, so answers are scoped to what the scene shows and say so.
Attachments
Drop files directly into the conversation. The assistant runs the same parsers as the upload pipeline: TLE sets, CSV, JSON, NDJSON. It previews what it found, imports on confirmation, and the resulting assets carry chat provenance in the registry. "Here is our ground station list, compare contact time across them" is one message.
Watching it work
Answers stream, and so does the work behind them:
- Tool-call trace. Each tool the assistant invokes appears in the transcript as it runs, with its arguments and result summary. You watch it select the link, run the trend query, and build the chart, in order, rather than receiving an unexplained conclusion.
- Insight cards. Analytical answers land as structured cards stating the number, the scope (which assets, which filters), the exact window and coverage, and the assumptions used. Cards are the audit trail: they are what gets captured into reports, and a card's numbers can be reproduced by re-running the same tools at the same sim time.
- Globe emphasis. When an answer names an asset, the asset pulses on the globe. The scene is part of the answer.
Stop and cancel
A running turn can be stopped mid-stream. Cancellation is clean: completed tool effects (a selection, a seek) remain, since you watched them happen, and the transcript marks the turn as stopped rather than presenting a partial answer as finished. Long analyses never hold the console hostage.
Connection modes
The assistant reports its own connectivity honestly in Settings:
| Mode | Meaning |
|---|---|
live | Production chat service; full reasoning and tools |
local | Development server with a local chat backend; same behavior, local keys |
offline | No chat backend reachable; built-in local intents only |
Offline mode still works for a useful subset: deterministic fleet metrics, chart requests, and common commands run entirely in the browser against the scene, with no model in the loop. The transcript labels which path answered.
Limits and plans
Guests get 10 assistant messages to evaluate with; signed-in plans are unlimited and select the reasoning model and prediction fidelity tiers, per Permissions and tenancy. When a guest hits the limit, the composer says so plainly and keeps the rest of the console fully usable.
Common workflows
| Task | Say |
|---|---|
| Scoped diagnosis | Select a link, then "why did SNR drop here in the last hour" |
| Cross-asset comparison | "@Oslo-GW vs @Svalbard-GW contact minutes today" |
| Data ingestion | Drop a TLE file, "import these and show them next to the fleet" |
| Scene control | "seek to 12:30 UTC", "show Starlink", "frame Europe" |
| Scenario work | "simulate losing both Norwegian gateways" |
| Reporting | "capture this analysis into the weekly capacity report" |
Example scenarios
An incident, hands off the mouse. During a degradation, an operator asks "which assets degraded in the last 30 minutes and what do they share". The trace shows the posture query, a breakdown by node type, and a trend over the window; the insight card names the affected gateway region and the exact window. The operator stops the follow-up chart mid-generation once the cause is obvious and pastes the card into the incident channel.
A planning session from a file. A planner drops a candidate-site CSV into chat, has the sites imported as synthetic stations, and asks for a ranked contact-time comparison against the propagated fleet. Every number arrives with its window, mask, and clear-sky assumption stated, and the whole exchange is captured into a gateway proposal report.