docs · concepts

Escalation by default

concept

When sources are silent, VeriCite escalates to a human instead of guessing.

VeriCite is built to know when to stop. When your approved sources do not support a confident answer, the Assistant does not improvise — it declines, and where you have configured it, it hands the question to a human. Escalating on silence rather than guessing is the behaviour that keeps a wrong answer from reaching a user, and it is assembled from three distinct safety layers working in sequence: the Scope Gate, confidence bands, and the Fallback policy.

Knowing when to stop is a product feature — A general chatbot fills silence with a plausible guess. VeriCite treats "I don't have a confident answer for that" — and a clean hand-off to a person — as the correct outcome, not a failure. Every refusal and hand-off is grounded in whether your sources actually support the answer.

What it is

"Escalate by default" is the union of three independent checks that decide what happens to a question the Assistant cannot safely answer from your Corpora. They fire at different points in the pipeline and are not interchangeable — the same user message can be deflected by one and never reach the others.

Scope Gatepre-retrieval

Decides whether an inbound question is *in scope* for your Assistant **before** retrieval runs, deflecting off-topic queries. If a question is out of scope, the Assistant declines early — it never searches your Corpora for something you never meant it to answer.

Guardrailscontent safety

Input and output content filters — PII, injection, toxicity, hallucination. Guardrails are about the *safety of the content* on the way in and the way out, not about whether your sources support the answer. They are a separate concern from escalation.

Fallback policylow retrieval confidence

What the Assistant does when retrieval confidence is *too low* to answer well — refuse, hand off to a human, or attempt a best-effort answer. This is the layer that turns "my sources are silent" into a refusal or an escalation.

Three concepts, not three names for one thing — The Scope Gate (off-topic), Guardrails (unsafe content), and the Fallback policy (low retrieval confidence) are distinct. An off-topic question is deflected by the Scope Gate and never reaches the Fallback policy; a well-scoped question with no supporting sources sails past the Scope Gate and is caught by the Fallback policy instead.

Why it matters

In a trusted-answer setting, a confident wrong answer is more damaging than an honest refusal. Escalate-by-default means a user at your institution either gets an answer your approved sources actually back — with Citations — or they get told plainly that the Assistant cannot answer, optionally routed to a person who can. You decide where that line sits, and a declined question becomes a signal you can act on rather than a silent failure.

How it relates to other concepts

Confidence bands

Every answer the Assistant produces carries a confidence band. The widget renders one of three levels, and the band is what the Fallback policy reasons about when it decides whether retrieval was strong enough to answer at all.

highband

Retrieval strongly supports the answer; the Assistant answers with Citations.

mediumband

Partial support; the answer is shown, often with a softer framing.

lowband

Weak or no support. This is the territory where the Fallback policy takes over — refuse, hand off, or best-effort.

The Fallback policy and the shipped controls

The Fallback policy is configured on the **Fallback** screen. Today that screen is a single, simple inline form with three fields: **Minimum retrieval score (0.0–1.0)** — below this score the Assistant refuses; **Refusal copy** — the message shown when it declines; and **Escalation email** — where a declined question can be routed.

The shipped Fallback screen is the simpler form — A richer Fallback design exists in the codebase — a behaviour enum (refuse / human_handoff / best_effort) plus an escalation queue — but it is not wired into the live screen. What ships and saves is the inline three-field form above. Describe the behaviour you can actually configure: a retrieval-score threshold, refusal copy, and an escalation email.

Hand-off — routing a question to a human

Where escalation means "send this to a person," the routing is configured per Assistant alongside its System prompts. You set per-channel hand-off targets, and each target carries a UX mode that controls how visible the hand-off is to the user.

silentUX mode

The hand-off happens behind the scenes; the user is not told.

notifyUX mode

The user is shown a message that their question has been routed onward.

confirmUX mode

The user is asked to confirm before the hand-off proceeds.

Limitations to know

Two parts of the escalation story are incomplete in the shipped console, and the documentation is honest about both so you do not rely on a control that is not yet live.

The per-Assistant **Safety** panel is a stub. Its live data path is a no-op, so it renders fixture and empty values rather than your real safety activity, and its escalation tracking is labelled "coming soon." Do not treat the numbers or escalation rows on that panel as a working record of escalations.
The richer Fallback behaviour form (the refuse / hand-off / best-effort enum and escalation queue) is dead and unrendered. The live screen is the three-field inline form (Minimum retrieval score, Refusal copy, Escalation email).

Escalation tracking is not live yet — Configuring an Escalation email on the Fallback screen sets where declined questions can be routed, but the per-Assistant Safety panel that would let you watch escalations land is still a Tier-C stub. To see questions the Assistant could not answer, use the Review queue, where flagged items collect.

Citations — how every supported answer carries references back to your approved sources.
Review queue — where flagged questions (unanswered, or thumbs-down) collect for a human to resolve.

conceptCitationsEvery answer carries citations to the approved sources it was generated from.

featureReviewReview answers, handle escalations, and reply as a human expert.

Was this page helpful?