FAQ

Is it safe to ask APEX about production data?

Yes — within limits. APEX is HIPAA fail-closed: it queries live systems but only returns aggregates from BigQuery and never posts raw clinical text, identifiers, patient names, emails, or Firebase UIDs. External users are blocked entirely.

Why did APEX ignore my message?

APEX only responds to a real @APEX mention. Quoting the word "APEX", editing an old message, or messaging from an external account won't trigger it.

What does the confidence number mean?

Every answer is graded by evidence tier (A: 0.95, B: 0.85, C: 0.75, D: 0.60). If overall confidence is below the 0.72 triage threshold, APEX flags the answer for a human instead of asserting it.

Does APEX remember past conversations?

No — it's completely stateless. Each interaction is processed in isolation, and pods restart every few hours, wiping in-memory state. The only "state" lives in Slack itself (the emoji reactions). A persistent memory / knowledge-graph layer is a proposed future direction, not something shipped today.

Will APEX open a PR or file a ticket automatically?

Only when you explicitly ask. By default it answers questions; creating PRs, tickets, or docs is an opt-in action.

How is APEX related to Frankie?

Frankie is a coding agent built on APEX — it reuses the same dispatcher → BigQuery queue → worker pod runtime, and inherits APEX's directive, HIPAA rules, and live-source connections. Sentinel (deployment monitoring) and Ascend (RFP workflows) run on the same pattern. See the Frankie docs.

How does APEX scale, and what happens mid-answer during a deploy?

APEX runs ~10 Socket Mode pods plus 1–5 watcher pods on an HPA, handling ~100 concurrent agents. Rolling updates use maxUnavailable: 0 / maxSurge: 2 for zero downtime, and a 62-minute termination grace period lets in-flight answers drain before a pod exits.

Where's the source of truth for behavior?

agents/DIRECTIVE.md in abridgeai/apex — 19 investigation steps, evidence tiers, HIPAA/PHI rules, environment awareness, Slack attribution, response length, and safety.