Astra (AI assistant)

Astra is Orbtrace's built-in AI assistant for on-call work. It's not a chatbot bolted on the side — it can actually query your telemetry, fetch a trace, run a root-cause analysis, and draft an alert rule, then show you the result inline. It opens anywhere with a keyboard shortcut and remembers the screen you're on.

It needs an AI provider configured (Admin → AI) and is a licensed feature — on the free COMMUNITY edition it's unavailable. See Licensing.

Opening it

Press ⌘J (Ctrl-J) anywhere in the app. Astra slides up as a floating panel that persists across navigation — you can keep a conversation going while you move between screens, and the panel won't dismiss when you click outside it (so a long, streaming answer isn't lost). The context strip shows which screen and time window Astra currently "sees".

What it can do

Astra has a set of tools it calls on your behalf; you ask in plain language and it picks the right one. Among them:

When Astra calls a tool, the result appears as a preview chip right in the chat — a log search renders as a results table, a trace as a mini-waterfall, a metric as a sparkline you can hover, an RCA as its diagnosis card, a drafted alert as an editable preview. You're never looking at raw JSON.

"What I noticed" — proactive watching

Astra doesn't only answer when asked. It watches your open anomalies and surfaces them in a "What I noticed" section: error spikes, latency regressions, and rate spikes, each with a severity (from the anomaly's z-score), a one-line baseline summary, and a pivot link to the affected service and time window. The floating button carries a badge counting the warning/critical items, and clicking one pre-seeds a chat turn with that context so you can dig in immediately. Dismiss anything that's noise.

Controls and footer

• Stop — a streaming answer can be cancelled mid-flight (the rose Stop button); the partial reply is frozen so you keep what arrived.
• Onboarding — on an empty conversation Astra offers a few starter prompts wired to your current screen.
• Cost footer — the composer shows running token totals for the session, with a ~ prefix when any number is an estimate. Per-tool, per-operator budgets can cap usage (set in configuration); when a tool's hourly budget is spent, Astra declines that tool until it resets.
• Sessions — past conversations are browsable at Astra → Sessions; open one to replay the full transcript, or delete it.

Guardrails worth knowing

• Citations, not hallucinations. RCA answers must cite the span IDs they reasoned from, so you can click through and verify.
• PII redaction. If you enable it (Admin → Astra PII), Astra scrubs emails, SSNs, phone numbers, credit-card and key-like strings from your text and the active query before anything reaches the LLM or the saved history. It's off by default; turn it on if prompts may carry regulated data. The model sees [REDACTED:…] placeholders, so it keeps structural awareness without seeing the value.
• Audit trail. Every chat turn is written to the audit log with the principal, provider, model, latency, and per-pattern PII counts — never the redacted values themselves.
• Provider choice. Astra routes through the same provider you pick in Admin → AI, so air-gapped deployments run it fully offline on Ollama.

Where it fits

Astra is the fast path for the questions you'd otherwise click through five screens to answer. For the deep, structured surfaces it pulls from, see Incidents / RCA, Time-Travel Replay, and the Alerts console.