> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gurubase.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Trust Layer

> Citations on every line. Refusals you can audit. The product surface that makes a Guru hallucination-resistant.

Generic LLMs guess when they should stop. The Gurubase **Trust Layer** wraps every answer in three product guarantees:

1. **Citations** point back to the exact paragraph the answer came from.
2. **A trust score** quantifies how well the retrieved context covered the question.
3. **Refusals** carry the policy attached when scope is broken or context is too weak.

The result is an AI agent your CISO, your auditor, and your end users can all trust.

<Note>
  For the deeper, multi-layer mechanism behind this guarantee (smart retrieval,
  multi-pass evaluation, source-grounded generation), read [Preventing AI
  Hallucinations](/guides/preventing-hallucinations). This page covers the
  user-visible product surface.
</Note>

## Three Product Guarantees

<CardGroup cols={3}>
  <Card title="Cited" icon="quote">
    Every answer includes source pills linking back to the original document, ticket, page, or paragraph.
  </Card>

  <Card title="Scored" icon="gauge">
    A trust score on every reply reflects retrieval quality and context coverage. Low scores show as warnings; very low scores trigger refusal.
  </Card>

  <Card title="Refused" icon="ban">
    When context is weak or out of scope, the Guru refuses and surfaces the policy reason instead of guessing.
  </Card>
</CardGroup>

## How a Trusted Answer Looks

Every answer renders three signals together:

* **Body**: the answer itself, grounded in the retrieved source paragraphs.
* **Source pills**: a list of cited documents (e.g. `hr-policy.pdf · p.7`, `SLA-2024-07`).
* **Trust badge**: a score between 0 and 1. High (0.85+) means strong retrieval coverage. Low scores warn the user; below the threshold the Guru refuses.

When context is too thin, the Guru produces a **refusal** instead of an answer: a short message that explains *why* it cannot answer (out of scope, no source, conflicting sources) and points the user to a human path. Every refusal lands in the audit log next to a normal answer, so a reviewer can see what the Guru did *not* say and why.

## What Triggers a Refusal

<CardGroup cols={2}>
  <Card title="Weak retrieval" icon="zoom-out">
    The Guru could not find source paragraphs above the relevance threshold for the question.
  </Card>

  <Card title="Out-of-scope question" icon="signpost">
    The question is outside the configured topics or sources of this Guru.
  </Card>

  <Card title="Conflicting sources" icon="triangle-alert">
    Multiple sources disagree and the Guru cannot pick a confident answer.
  </Card>

  <Card title="Policy-attached scope" icon="scale">
    A configured rule (PII, compliance, RBAC) blocks the question from this user or channel.
  </Card>
</CardGroup>

## Trust Score Threshold

You can tune the trust score threshold per Guru. Lower thresholds answer more questions but with weaker grounding. Higher thresholds refuse more aggressively. Regulated teams (finance, healthcare, legal) usually run with a stricter threshold; community support and developer docs typically run looser. [Contact us](https://gurubase.io/contact/) to configure the threshold for your Guru.

## Source-Level RBAC and PII Masking

The Trust Layer also enforces what each user is allowed to see:

* **Source-level RBAC**: tag sources by audience (e.g. internal-only, finance-only) and the Guru will refuse to cite them outside that audience.
* **[PII masking](/guides/pii-masking)**: auto-detect and mask emails, phone numbers, and other personal data before they reach the answer or the audit log.
* **[Audit logs](/guides/audit-logs)**: every answer and every refusal is logged with sources, trust score, and the user who asked.

## Where to Go Next

<CardGroup cols={2}>
  <Card title="Preventing AI Hallucinations" icon="shield-check" href="/guides/preventing-hallucinations">
    The seven-layer mechanism behind citations, scoring, and refusals
  </Card>

  <Card title="PII Masking" icon="shield-user" href="/guides/pii-masking">
    Auto-mask personal data before it leaves the Guru
  </Card>

  <Card title="Audit Logs" icon="clipboard-list" href="/guides/audit-logs">
    Track every answer, refusal, and source on a regulated-deployment timeline
  </Card>

  <Card title="Triage" icon="route" href="/guides/triage">
    Clarify ambiguous questions before the Guru ever cites a source
  </Card>
</CardGroup>