# Re-Rank Min Coverage %: a floor on how much an answer is backed

> Re-Rank Min Coverage % drops answers that fall below a coverage threshold — a hard floor beneath which a response simply won't ship.

**Category:** Hallucination Prevention
**Author:** NeuralSeek Team · **Published:** June 9, 2026
**Canonical:** https://neuralseek.ai/ai-grounded/re-rank-min-coverage
**Section index:** https://neuralseek.ai/ai-grounded

Even after re-ranking has surfaced the best evidence and coverage weighting has measured it honestly, some answers are simply too thinly supported to trust. Re-Rank Min Coverage % is the control that draws the line. It sets a hard floor: if the share of an answer actually backed by source falls below your threshold, the answer is dropped rather than delivered. It's the platform's way of encoding the judgment that, sometimes, saying nothing is the right answer.

## What it actually does

The guardrail computes how much of the candidate answer is covered by the re-ranked sources and compares that figure to a configured minimum percentage. Answers that clear the floor proceed; answers below it don't make it through. Instead of being presented as grounded, they're withheld or routed to a fallback path — a human reviewer, a clarifying question, or an alternate workflow — so the user never receives a response the system couldn't adequately back.

## Why business teams care

This is the 'better to say nothing' control, and in high-stakes settings it is invaluable. An unsupported answer is frequently worse than a graceful 'I don't have enough information to answer that,' because the unsupported answer carries the full authority of the assistant behind a claim it can't defend. The coverage floor encodes that judgment as an enforceable rule rather than leaving it to the model's discretion, which is exactly where you don't want this decision living.

## How to tune it in practice

Set the floor according to the cost of a wrong answer in each workflow. Where that cost is high — regulated advice, financial figures, medical or legal content — raise the floor so the assistant declines unless it's genuinely well-supported. Where partial answers still add value and a human is in the loop, lower it to favor coverage. Track the decline rate: if it's climbing on questions your knowledge base should answer, that's a signal to revisit either the floor or the content behind it.

## Common failure modes it prevents

The failure it directly prevents is the 'thin answer shipped as fact' — a response grounded in just enough source to score above zero but nowhere near enough to be reliable. It also prevents the slow erosion of trust that happens when users repeatedly catch the assistant overreaching on weak evidence. A well-set floor means that when the assistant does answer, the answer is substantively backed, which is what makes its confidence credible.

## Where it fits in the stack

Re-Rank Min Coverage % sits at the end of the grounding pipeline, acting on the coverage measurements produced upstream by Total Coverage Weight and the re-ranking pass. It's one of the final gates before an answer reaches the confidence layer and delivery — the last hard checkpoint that can stop a poorly-supported answer from ever being seen. Its reliability depends on the coverage measurement beneath it being honest, which is why it pairs so tightly with Total Coverage Weight.

## Tuned to risk tolerance

Raise the floor where the cost of a wrong answer is high; lower it where partial answers still help and the downside is contained. Either way, the threshold makes your organization's risk appetite explicit, consistent, and auditable — applied identically to every request rather than varying with the model's mood.

> Knowing when not to answer is as much a feature as answering well.

## The takeaway

Re-Rank Min Coverage % puts a hard floor under grounding — thinly supported answers are dropped, not delivered — turning the discipline of restraint into an enforceable guarantee.

---

From NeuralSeek's AI Grounded — practical, web-verified guidance on building governed, grounded enterprise AI. NeuralSeek is the model-agnostic, governed AI platform you own: any LLM (swap with no rebuild), your data in your own tenant (cloud or on-prem), 118 guardrails enforced before any action, one container that runs anywhere.